Radar Interferometry

Remote Sensing and Digital Image Processing

VOLUME 2

Series Editor:

Freek van der Meer, International Institute for Aerospace Survey and
Earth Sciences, ITC, Division of Geological Survey, Enschede, The Netherlands and
Department of Applied Earth Sciences, Delft University of Technology, The Netherlands

Editorial Advisory Board:

Michael Abrams, NASA Jet Propulsion Laboratory, Pasadena, CA, U.S.A.

Paul Curran, University of Southampton, Department of Geography,
Southampton, U.K.
Arnold Dekker, CSIRO, Land and Water Division, Canberra, Australia
Steven de Jong, Wageningen University and Research Center,
Center for Geoinformation, Wageningen, The Netherlands
Michael Schaepman, ETH, Zurich, Switzerland
RADAR INTERFEROMETRY
Data Interpretation and Error Analysis

by

RAMON F. HANSSEN
Delft University of Technology, The Netherlands

KLUWER ACADEMIC PUBLISHERS

NEW YORK, BOSTON, DORDRECHT, LONDON, MOSCOW
eBook ISBN: 0-306-47633-9
Print ISBN: 0-7923-6945-9

©2002 Kluwer Academic Publishers

New York, Boston, Dordrecht, London, Moscow

Print ©2001 Kluwer Academic Publishers

Dordrecht

All rights reserved

No part of this eBook may be reproduced or transmitted in any form or by any means, electronic,
mechanical, recording, or otherwise, without written consent from the Publisher

Created in the United States of America

Visit Kluwer Online at: http://kluweronline.com

and Kluwer's eBookstore at: http://ebooks.kluweronline.com
 Contents

Preface ix

Summary xiii

Nomenclature xv

1 Introduction 1
1.1 Motivation 1
1.2 Background 1
1.3 Problem formulation, research objectives and limitations 3
1.4 Research methodology 6
1.5 Outline 7

2 Radar system theory and interferometric processing 9

2.1 Radar history and developments 9
2.2 Sensor characteristics 22
2.3 Image formation 23
2.4 SAR interferometry 34
2.5 Interferometric processing overview 42

3 Functional model for radar interferometry 61

3.1 Gauss-Markoff model definition 61
3.2 Topography estimation 66
3.3 Deformation mapping 69
3.4 Atmospheric mapping 70
3.5 Generic model interferogram stacks 72
vi Contents

4 Stochastic model for radar interferometry 81

4.1 Theory 82
4.2 Single-point observation statistics 87
4.3 Coherence and SNR 96
4.4 Sources of decorrelation 98
4.5 Integer phase ambiguities 111
4.6 Influence and modeling of orbital errors 113
4.7 Atmospheric signal: turbulence 130
4.8 Atmospheric signal: stratification 148
4.9 Error propagation and joint stochastic model 154
4.10 Conclusions 159

5 Data analysis and interpretation for deformation monitoring 161

5.1 Decomposition of the displacement vector 162
5.2 Corner reflector experiments 163
5.3 Groningen coherence estimation 173
5.4 Cerro Prieto geothermal field 175
5.5 Izmit earthquake 190
5.6 Conclusions 194

6 Atmospheric monitoring 197

6.1 Theory 199
6.2 Refractivity sensitivity analysis 204
6.3 Ionospheric influence 209
6.4 Water vapor mapping 215
6.5 Combining amplitude and phase information 220
6.6 Validation radar interferometry using Meteosat radiometry 226
6.7 Spatial moisture distribution during CLARA 96 237
6.8 Conclusions 244

7 Conclusions and recommendations 247

7.1 Contributions of this research 247
7.2 Recommendations 251

A Comparison neutral delay GPS and InSAR 255

A.1 Spatial networks 256
 Contents vii

A.2 Single-station time-series 264

B Structure function and power spectrum 271

B.1 Structure function for refractivity and signal delay 271
B.2 Structure function and power-law spectrum 273

Bibliography 277

About the Author 299

Index 301
 Preface

This book is the product of five and a half years of research dedicated to the under-
standing of radar interferometry, a relatively new space-geodetic technique for mea-
suring the earth’s topography and its deformation. The main reason for undertaking
this work, early 1995, was the fact that this technique proved to be extremely useful
for wide-scale, fine-resolution deformation measurements. Especially the interfer-
ometric products from the ERS-1 satellite provided beautiful first results—several
interferometric images appeared as highlights on the cover of journals such as Nature
and Science. Accuracies of a few millimeters in the radar line of sight were claimed
in semi-continuous image data acquired globally, irrespective of cloud cover or solar
illumination. Unfortunately, because of the relative lack of supportive observations
at these resolutions and accuracies, validation of the precision and reliability of the
results remained an issue of concern.
From a geodetic point of view, several survey techniques are commonly available to
measure a specific geophysical phenomenon. To make an optimal choice between
these techniques it is important to have a uniform and quantitative approach for
describing the errors and how these errors propagate to the estimated parameters.
In this context, the research described in this book was initiated. It describes issues
involved with different types of errors, induced by the sensor, the data processing,
satellite positioning accuracy, atmospheric propagation, and scattering characteris-
tics. Nevertheless, as the first item in the subtitle “Data Interpretation and Error
Analysis” suggests, data interpretation is not always straightforward. Especially
when the interferometric data consist of a superposition of topography, surface de-
formation, and atmospheric signal, it is important to recognize the characteristics
of these signals to make a correct interpretation of the data. In this work, I hope to
contribute to improved error analysis and data interpretation for radar interferom-
etry.
This book owes significantly to the people I had the pleasure to work with during
the past several years. First of all, I would like to thank Roland Klees for making
it all possible, for supporting me to work abroad for such a long time, and for his
supervision. My room mates Bert “Mr. Doris” Kampes and Stefania Usai provided
a nice working environment and enough food for lengthy discussions.
I learned a lot from the MSc-students, whom I had the pleasure to advise during
the last couple of years. Jaron Samson, Yvonne Dierikx-Platschorre, Ronald Stolk,
x Preface

Claartje van Koppen, and Rens Swart, your work has contributed significantly to
the results described in this book. Appendix A is based on ideas to combine GPS
data with SAR interferograms for the correction or estimation of atmospheric error.
The GPS data processing and analysis in this appendix was performed by Ronald
Stolk and André van der Hoeven, supported by Hans van der Marel and Boudewijn
Ambrosius.
The meteorological interpretation of the radar interferograms was only possible due
to the close collaboration with several meteorologists. At the Royal Netherlands Me-
teorological Institute (KNMI) Sylvia Barlag, Frans Debie, Arnout Feijt, and many
others participated in these analyses. Arnout Feijt developed the Meteosat water
vapor channel parameterization using the GPS time series for one of the case studies
of chapter 6, which enabled the Meteosat-InSAR validation. I am highly indebted
to Tammy Weckwerth at the National Center for Atmospheric Research in Boulder,
Colorado, who devoted much of her time to the interferogram interpretation lead-
ing to our Science paper. Our discussions in Boulder, as well as our email-battles
improved my understanding of meteorology significantly. Stick Ware, thanks for
making the link! I was introduced into the wonders of SAR amplitude imagery by
Susanne Lehner, Ad Stoffelen, and Ilona Weinreich. The conformity between the ob-
served wind patterns and the water vapor distribution provided a consistent support
for the interferogram interpretation.
During the years we performed several common research projects with the Physics
and Electronics Laboratory of TNO. Marco van der Kooij (now at Atlantis Sci-
entific) and Erik van Halsema introduced me to interferometry and the Groningen
land subsidence project. In a later stage, Jos Groot and Roel van Bree had a sig-
nificant influence on my understanding of airborne interferometry (PHARUS) and
in the Tianjin land subsidence project with EARS. At the Survey Department of
Rijkswaterstaat, Erik de Min, Jur Vogelzang, and Yvonne Dierikx actively pursued
the “practical relevance” issue, which kept me on track from time to time.
During my stay at the Institut für Navigation at Stuttgart University, Karl-Heinz
Thiel, Jürgen Schmidt, and dr. Wu enabled me to obtain valuable experience in
interferometric data processing. My roots in the understanding of radar lie at DLR-
DFD in Oberpfaffenhofen. I am grateful to Richard Bamler, Michael Eineder, Nico
Adam, and many others for giving me the opportunity to work with them, which
was pleasing in many ways. At Stanford University I owe much to Howard Zebker
for his support and hospitality. Discussions with him, Paul Segall, Sjonni Jónsson,
and Weber Hoen were always enlightening. Now at the University of Hawaii, Falk
Amelung was my tutor in geophysics. Falk, aloha for the Cerro Prieto Friday after-
noon analysis. I am indebted to Ewa Glowacka at CICESE, Ensenada, and CFE,
Mexico for their support in the Cerro Prieto study. David Sandwell is gratefully
acknowledged for inviting me to visit Scripps Institution of Oceanography, which
was a great experience.
Many people contributed to the GISARE field experiment, including the meteorolog-
ical, the leveling, the GPS, and the SAR processing part. Installing corner reflectors
in frozen ground at –15°C requires special skills, we had a great team. Although
the GPS-SAR experiment during the solar eclipse in 1999 failed to produce an ob-
 Preface xi

servable ionospheric signal in the interferograms, it provided good experience thanks

to all people contributing.
Special thanks go to Herman Russchenberg for his droplets, Michel Decré for fig. 2.1A,
Riccardo Riva and Bert Vermeersen for the Izmit earthquake analysis, and Remko
Scharroo for his precise satellite orbits.
The manuscript of this book greatly benefited from the meticulous reading of my
defense committee: Richard Bamler, Philipp Hartl, David Sandwell, Tammy Weck-
werth, Jacob Fokkema, Peter Teunissen, and Roland Klees, thank you for all your
critical comments and valuable improvements.
Freek van der Meer, series editor, and Petra van Steenbergen at Kluwer Academic
Publishers are gratefully acknowledged for making the monograph publication of
this work possible in the Remote Sensing and Digital Imaging Processing series.
This book could not have been completed without the research funds of the Cor-
nelis Lely Stichting, the Survey Department of Rijkswaterstaat, the European Space
Agency, the Fulbright program, the Universiteitsfonds, and Shell International Pro-
duction and Exploration B.V. All ERS SAR data were kindly provided by ESA.
Meteorological data were made available by the Royal Netherlands Meteorological
Institute (KNMI). Maps and visualizations have been implemented in GMT (Wessel
and Smith, 1998) and in Matlab™.
Finally, I would like to thank Ardis for her understanding and support and for all
the time I stole from her. I’m excited about the future with the three of us.

Ramon Hanssen
Delft, December 2000
 Summary

Within a decade, imaging radar interferometry has matured to a widely used geodetic
technique for measuring the topography and deformation of the earth. In particu-
lar the analysis and interpretation of the interferometric data requires a thorough
understanding of the principles of the technique, the (potential) error sources, and
the error propagation. This book reviews the basic concepts of radar, imaging
radar, and radar interferometry, and revisits the processing procedure for obtaining
interferometric products such as a digital elevation model or a deformation map.
It describes spaceborne repeat-pass radar interferometry using a linear or Gauss-
Markoff model formulation, which relates the interferometric observations to the
unknown geophysical parameters. The stochastic part of the model describes the
dispersion of the observations in terms of variances and covariances. Especially the
influence of spatially correlated errors as induced by the satellite orbits and by atmo-
spheric path delay are discussed. Mathematical models are presented that describe
the spatial variability in the interferometric phase due to turbulent mixing of at-
mospheric refractivity and due to vertical atmospheric stratification. Using 52 SAR
acquisitions, a systematic inventory of the characteristics of atmospheric signal in
the radar interferograms is performed, using complementary meteorological data for
the interpretation. Scaling characteristics are observed, which can be conveniently
used to describe the power spectrum and covariance function of the atmospheric
signal. The final variance-covariance matrix for the radar interferometric data is
presented, including these spatially varying error sources. A number of case studies
on deformation monitoring, such as land subsidence, earthquake deformation, and
artificial reflector movement serve as examples of the application of interferometry
and its error sources. The feasibility of the technique for practical geodetic appli-
cations is evaluated in relation to the geophysical phenomena of interest, yielding
rules-of-thumb for its utilization. Finally, a novel application of interferometry for
atmospheric studies, termed Interferometric Radar Meteorology, is presented and
discussed. Maps of the vertically integrated water vapor distribution during the
radar acquisitions can be obtained with a fine spatial resolution and a high accu-
racy. Several demonstration studies of this meteorological application are presented.
 Nomenclature

List of acronyms
1D One-Dimensional
2D Two-Dimensional
ALD Azimuth Look Direction
ALOS Advanced Land Observing Satellite
AMI Active Microwave Instrument
ASI Agenzia Spaziale Italiana (Italian Space Agency)
AVHRR Advanced Very High Resolution Radiometer
BTTB Block-Toeplitz Toeplitz-Block
CFE Comisión Federal de Electricidad, Mexico
CLARA’96 Clouds and Radiation experiment 1996
CMOD4 C-band Model 4 (empirically derived model function to relate normal-
ized radar cross-section with wind speed and direction)
CNES Centre National D’Etudes Spatiales
CPGF Cerro Prieto Geothermal Field
CSA Canadian Space Agency
CTRS Conventional Terrestrial Reference System
DEM Digital Elevation Model
DEOS Delft Institute for Earth-Oriented Space Research
DFD Deutsches Fernerkundungsdatenzentrum
(German Remote Sensing Data Center)
DIAL Differential Absorption Lidar
DLR Deutsches Zentrum für Luft- und Raumfahrt e.V.
(German Aerospace Center)
DPWV Differential Precipitable Water Vapor
DUT Delft University of Technology
ERS European Remote Sensing Satellite
ESA European Space Agency
fBm Fractional Brownian Motion
FFA Far Field Approximation
FFT Fast Fourier Transform
FM Frequency Modulated
GISARE Groningen Interferometric SAR Experiment
GNSS Global Navigation Satellite Systems
GPS Global Positioning System
IEEE Institute of Electrical and Electronics Engineers
IERS International Earth Rotation Service
IGS International GPS Service for Geodynamics
InSAR Interferometric SAR / SAR interferometry
IPW Integrated Precipitable Water
I/Q In-Phase (real), Quadrature (imaginary)
IRM Interferometric Radar Meteorology
xvi List of symbols

ITRF IERS Terrestrial Reference Frame

JERS Japanese Earth-Resources Satellite
JPL Jet Propulsion Laboratory
LOS Line of Sight
LT Local Time
MCF Minimal Cost Flow
MIT Massachusetts Institute of Technology
MSC Mesoscale Shallow Convection
NAM Nederlandse Aardolie Maatschappij B.V.
NAP Normaal Amsterdams Peil (Dutch vertical reference datum)
NASA National Aeronautics and Space Administration
NASDA National Aeronautics and Space Development Agency (Japan)
NDVI Normalized Differential Vegetation Index
NOAA National Oceanic and Atmospheric Administration
PAF Processing and Archiving Center
PDF Probability Density function
PLW Precipitable Liquid Water
PRARE Precise Range and Rate Equipment
PRF Pulse Repetition Frequency
PWV Precipitable Water Vapor
radar Radio detection and ranging
RAR Real Aperture Radar
RCS Radar Cross Section
RH Relative Humidity
RMS Root Mean Square
RRF Range Reference Function
SAR Synthetic Aperture Radar
s/c Spacecraft
SIR Shuttle Imaging Radar, Spaceborne Imaging Radar
SLAR Side Looking Airborne Radar
SLC Single-Look Complex
SLR Satellite Laser Ranging / (Side Looking Radar)
SNR Signal-to-Noise Ratio
SPOT System Probatoire d’Observation de la Terre
SRTM Shuttle Radar Topography Mission
SWST Sampling Window Start Time
TEC Total Electron Content
TECU Total Electron Content Unit
TID Traveling Ionospheric Disturbance
UTC Universal Coordinated Time
VLBI Very Long Baseline Interferometry
VTEC Vertical Total Electron Content
WGS84 World Geodetic Survey 1984
WV Water Vapor
X-SAR X-band Synthetic Aperture Radar

List of symbols
Initialization value, usually determined for reference surface
Design matrix
Part of the design matrix corresponding to observation
Baseline
Parallel baseline
Perpendicular baseline
Perpendicular baseline for range to resolution cell on the reference sur-
face
Critical (perpendicular) baseline
Horizontal baseline
 List of symbols xvii

Vertical baseline
Azimuth time bandwidth
Range bandwidth
Doppler bandwidth
Speed of light = 299792458 m/s
Structure function coefficient
Variance-covariance matrix of the estimated parameters
Variance-covariance matrix
Variance-covariance matrix of atmospheric delay
Antenna width (m)
Dispersion operator, second moment, or variance-covariance matrix
Fractal dimension for a signal
Slant deformation for observation
Structure function of parameter
Expectation operator, ensemble average, first moment, or mean
Partial pressure of water vapor (hPa)
Frequency
Carrier (center) frequency
Doppler frequency
Doppler centroid frequency
Pulse repetition frequency (PRF) (cycles/m)
Fringe frequency
Altitude of point
Height of the sub-ionospheric point (m)
Height ambiguity (m)
Altitude; Haussdorff measure
Topographic height for observation
Integrated Precipitable Water
Wavenumber
Effective number of looks
Antenna length (m)
GPS carrier phase frequencies (Ll=l.57542 GHz, L2=l.22760 GHz)
GPS ionospheric-free linear combination
Number of observations
Number of parameters (unknowns)
Electron content
Parallel error baseline
Perpendicular error baseline
Refractivity, Geoid height
Probability density function
Total atmospheric pressure (hPa)
Power spectrum of parameter
Real positive-semidefinite cofactor matrix
(Slant) range (m)
Slant range between reference (master) satellite and resolution cell (m)
Slant range between secundary (slave) satellite and resolution cell (m)
Chirp rate or Frequency slope
Slant atmospheric delay for pixel at (m)
Time (s)
Temperature (K)
Velocity (m/s)
Hermitian product
Velocity of spacecraft (m/s)
Integer ambiguity number
Liquid water content
Wrapping operator =
Radar footprint width in azimuth (3dB)
Vector of parameters
Best linear unbiased estimator of the unknown parameters
xviii List of symbols

State vector of spacecraft

Complex observation (phasor)
Amplitude spectrum of interferogram and are wavenumbers)
Amplitude spectrum of filtered interferogram
Baseline orientation angle; Spectral exponent
Error baseline orientation angle; Spectral index
Angular beamwidth in range direction (rad)
Angular beamwidth in azimuth direction (rad)
Coherence (complex)
Coherence reduction due to geometric (baseline) decorrelation
Coherence reduction due to Doppler centroid decorrelation
Coherence reduction due to temporal decorrelation
Coherence reduction due to volume decorrelation
Coherence reduction due to processing induced decorrelation
Coherence reduction due to thermal noise
Semi-variogram of parameter
Change in look angle
Focused azimuth resolution
Slant range resolution
Slant range difference
Slant range difference due to deformation
Vector of observation errors
Local terrain slope w.r.t. reference surface (positive towards the radar):

Look angle
Incidence angle, defined with respect to the global vertical at the scene
Local incidence angle, defined with respect to the local vertical at the
scene:
Reference phase (rad)

Radar wavelength
Geographic longitude in WGS84
Mean
Local earth radius (m); Horizontal distance between two resolution cells.
Autocorrelation function of parameter
a priori variance factor
“Sigma nought” or Radar Cross Section (RCS). Average power reflec-
tivity or scattering coefficient per unit surface area
Pulse length
Effective pulse length
Absolute interferometric phase, including reference phase
Wrapped interferometric phase, including reference phase
Geographic latitude in WGS84
Off-center beam angle in range direction (rad)
Off-center beam angle in azimuth direction (rad)
Squint angle (rad)
Phase noise (rad)
Absolute interferometric phase, reference phase subtracted
Wrapped interferometric phase, reference phase subtracted
Stochastic phase observation
Real stochastic vector of observations
Phase (rad)
Scattering contribution to the phase value (rad)
Ratio between of deformation pair and of topographic pair
Set of natural numbers
Set of real numbers
Set of integer numbers
Set of complex numbers
Fourier transform
 Chapter 1
Introduction

1.1 Motivation

The geodetic interpretation of measurements obtained by spaceborne repeat-pass

radar interferometry is limited by unmodeled error sources. The research reported
in this study is an attempt to situate the technique in a standard geodetic framework,
analyze the characteristics of the errors, and improve data interpretation.

1.2 Background

Advances in geodesy during the last couple of decades can be characterized by a

number of developments. First, there is a broadening of the spatial scales over
which measurements can be performed. The use of artificial satellites boosted the
definition and use of global reference systems and enabled the mapping of large
areas on earth. Optical remote sensing from spaceborne platforms triggered fast
developments in geodesy, especially using stereographic techniques, whereas point
positioning entered a new era using global navigation satellite systems (GNSS).
Recognition of the earth's dynamics characterizes a second development. General
acceptance of global plate tectonics does not only make every measurement a func-
tion of time, it also undermines the definition of reference systems (McCarthy, 1996).
On more local scales there is considerable interest in monitoring deformations in-
duced by earthquakes, volcanoes, glacier dynamics, post-glacial rebound, and an-
thropogenic deformation due to the exploration of mineral resources, geothermal
energy, and water. Better understanding of these processes requires regular repeat
measurements, high accuracy, and a fine spatial resolution. These requirements are
motivated by considerations of public safety and hazard monitoring, sea-level rise
and global warming, and controversial issues such as the underground storage of
nuclear material. In this context, geodetic observations serve as crucial source of in-
formation for geophysical interpretation. Regarding deformations caused by human
exploration legal and financial aspects play an important role (Taverne, 1993).
The accuracy of geodetic observations marks a third development. As a result of
the increased number of alternative techniques the quality, in terms of the precision
and reliability, needs to be balanced with the specific goals of the measurement
2 Chapter 1: Introduction

campaign. Important elements are financial constraints, accessibility of the area,

the duration of the measurement campaign, monumentation, and repeatability.
As a provider of geometric and gravimetric information, geodesy contributes to geo-
physics, geodynamics, and oceanography. Moreover, in a more indirect way space-
geodetic observations can contribute to atmospheric physics. The influence of atmo-
spheric refraction on the propagation velocity of electromagnetic waves is regarded
as one of the most important error sources for distance measurement, one of the
most elementary geodetic observables. Although space-geodetic techniques such as
very long baseline interferometry (VLBI), radar altimetry, and global navigation
systems such as GPS, were not designed for atmospheric studies, they contributed
considerably to the global understanding of the atmosphere. Currently, GPS obser-
vations are used on a routine basis by meteorologists, e.g., for the observation of
water vapor distributions.
Radar interferometry is perfectly embedded in this line of geodetic developments,
combining the characteristics of large-scale imaging and quantitative observations of
distances and angles. In terms of spatial scales, spaceborne radar interferometers are
able to collect observations distributed over the entire earth. Quantitative ranging
observations are currently collected over a swath width in the order of 100-500 km,
with a resolution of about 20 m. As opposed to optical sensors, the active radar
observations can be collected at nighttime, over areas with limited contrast, and
through cloud cover. Using a single-pass interferometric configuration, in which
two physical antenna’s are used to collect the radar echoes, accuracies for elevation
observations are in the order of 6 m for spaceborne sensors (Bamler and Hartl, 1998;
Werner et al., 1993; Adam et al., 1999; Jordan et al., 1996), which is comparable to
the optical techniques (Baltsavias, 1993; SPOT image company, 1995).
Nevertheless, the most spectacular application of radar interferometry is the obser-
vation of dynamic processes, using a repeat-pass configuration, in which two radar
images are acquired by a single antenna, revisiting the area after a specific time
interval. Coherent deformation of the earth’s surface within the radar swath can be
observed as a fraction of the radar wavelength, resulting in sub-cm ranging accura-
cies in the radar line-of-sight. The most stringent requirement for this application is
that the archive of radar data contains acquisitions prior to the deformation event.
The background of this study was formed by the recognition of these capabilities for
radar interferometry and the desire to consider the technique—which was conceived
in an electrical engineering environment—from a more geodetic point of view. Al-
though convincing demonstrations of the technique had been published, it was also
known, or at least expected, that there were limitations affecting the applicabil-
ity and accuracy. For example, for applications of spaceborne interferometry for
deformation studies in the Netherlands, previous studies identified main problems
due to the changing characteristics of the surface (temporal decorrelation) and at-
mospheric signal delay (van der Kooij et al., 1995). This study discusses these and
other error sources and the way the errors propagate to the final parameters of inter-
est. A conceptual approach is presented to regard the data using standard geodetic
methodology.
 1.3 Problem formulation, research objectives and limitations 3

The analysis of the atmospheric delay in the radar data has been one of the focus
points of this study. Especially the results of ESA’s ERS “tandem mission,” which
provided extremely coherent repeat-pass interferometric data, enabled an extensive
data set of “atmospheric” interferometric products. The close collaboration with
meteorologists lead to a remarkable result: apart from regarding atmospheric delay
as a source of error for deformation studies, it could be analyzed and interpreted
as a source of information for atmospheric studies. Another result from this new
application of interferometry is its use in improving stochastic models for the atmo-
spheric influence on other space-geodetic techniques such as GPS, radar altimetry,
or VLBI, especially regarding small-scale spatial disturbances.

1.3 Problem formulation, research objectives and limitations

The achievements of InSAR in terms of deformation monitoring and topographic

mapping have clearly demonstrated the unsurpassed capabilities of the technique.
Nevertheless, this does not necessarily prove its feasibility for any potential appli-
cation. Early applications of InSAR may be characterized as “technology driven,”
implying that its main goal was the demonstration of the capabilities. Only in re-
cent years, attention is shifting to “application driven” developments. The decision
whether or not to use InSAR for a specific application is not always straightforward.
There are a number of problems that influence this decision.

SNR The signal-to-noise ratio1 of the parameters under consideration is not always
known. Many demonstration studies have focused on areas where the defor-
mation or topography signal was dominant and error sources where marginal.
Without denying their fundamental importance, these studies may often be
characterized as rather opportunistic and ad-hoc. As application-oriented re-
search shifts to problems with a much more critical SNR, the demand on iden-
tifying and modeling the error sources increases. The main error sources that
influence the final results are (i) the contribution of atmospheric signal delay,
(ii) the interferometric decorrelation due to temporal and geometric scattering
characteristics, (in) the unknown phase ambiguity number, and (iv) errors in
supplementary information such as the reference satellite orbits and elevation
models. Models of the error sources and their propagation into the param-
eters of interest have been limited to single point statistics, largely ignoring
the quantitative description of covariance in the data, or even to a qualitative
interpretation of possible artifacts only. A full stochastic model, expressed
in terms of a variance-covariance matrix of the observations has not yet been
defined.

Data availability The regular acquisition of spaceborne SAR data is influenced

by power consumption considerations, downlink or data storage capabilities,
and possible conflicting user requirements on the satellite’s multi-instrument
payload. Furthermore, orbit maintenance requirements for topographic map-
1
Here the term SNR is used in an abstract sense, implying the relation between the parameters
of interest and all disturbing factors, whether deterministic or stochastic.
4 Chapter 1: Introduction

ping are fundamentally different from those for deformation mapping. Other
factors include the compatibility of polarization between the interferometric
pairs, the used incidence angles, and the radar frequency. These restrictions
limit the amount of useful interferometric data sets.
Ground truth Due to the unknown phase ambiguity number and the limited
knowledge of the satellite’s position, radar interferograms are essentially rel-
ative measurements. In order to relate these measurements to a reference
datum, a priori information is required. Control points or absolute deforma-
tion measurements from other geodetic techniques are not always available,
which influences the quality of the final results.
Surplus value The decision to use InSAR, instead of any other geodetic technique,
for a specific application will always be weighted based on characteristics such
as the spatial extent of the signal under study, the accessibility of the terrain,
desired accuracy and reliability, repeatability, processing speed, uniqueness of
the interpretation, and cost-effectiveness. This means that all these issues need
to be known before the decision to apply a specific technique is made.

In order to improve the availability and value of interferometric products, these

problems need to be considered.

Research objectives
The four categories that influence the decision to use InSAR for a specific geophysical
application (SNR, data availability, ground truth, and surplus value) comprise the
main limitations of the technique. It is beyond the scope of this study to comment
on instrument or spacecraft design and operation considerations that determine data
availability. Furthermore, when the radar interferometric observations and derived
parameters are clearly defined as being relative, the availability and use of ground
truth can be ignored. This study will focus on the first and the last item: the signal-
to-noise ratio and the surplus value. The general problem statement addressed in
this study is;

How can the interpretation and analysis of repeat-pass spaceborne radar

interferometric data be improved in a systematic way via a model-based
quantification of the error sources?

From this formulation of the problem, four specific research questions are derived
that give further direction to the treatment in this study.
1 How can the formation and the quality of a radar interferogram and its derived
interferometric products be parameterized?
The formation of a radar interferogram is a procedure that involves the raw radar
data acquisition and down-link from the spacecraft to the ground station on earth,
the evaluation of the data quality, and the data processing using a series of specific
operations. To enable a critical evaluation of the quality of each link in this chain,
it is evident that their physical and mathematical properties need to be described.
 1.3 Problem formulation, research objectives and limitations 5

The final result is a two-dimensional image representing some geophysical signal.

A parameterization of the quality of this signal is not always straightforward, as
errors with different characteristics will affect the result. Some errors will result
in a bias of all interferometric phase values, some will have varying characteristics
depending on the position in the image and some affect a single resolution cell only,
while others increase the variance of all resolution cells. Especially resolving the
superposed signals of topography, deformation, and atmosphere poses a significant
problem in this respect.
A conceptual approach would be to express a radar interferogram in a mathematical
model, consisting of a functional and a stochastic part. The functional model would
relate the basic phase observations to the parameters of interest, while the stochastic
model would describe the quality of the observations. As a next step, it could then
be investigated how the observations in several interferograms can be combined in
a single adjustment model.
2 Which error sources limit the quality of the interferogram and how does their
influence propagate in the interferometric processing?
In order to make quantitative statements on the quality of an interferometric prod-
uct, it is necessary to analyze the total chain from data acquisition via data process-
ing to interpretation, and identify the error sources in every step. Errors introduced
during the data acquisition will propagate into the final results, whereas specific
processing errors will only contribute to the final part of the data processing. Due
to the differences in the nature of the error sources, their effects can be very different
as well. Methods that are optimal to parameterize the statistics of every individual
step need to be developed.
3 To what extent do atmospheric refractivity variations affect the interferograms,
is it possible to determine this influence using additional measurements, and how
should it be modeled stochastically?
The effect of refractivity variations in the atmosphere is a known problem to space-
geodetic observations such as SLR, VLBI, radar altimetry, and GPS. Nevertheless,
its influence on radar interferometry has been ignored for several years, largely due
to the assumptions of the all-weather capability of radar, which has been one of the
main arguments for using radar instead of optical remote sensing from the begin-
ning. The problem is now recognized as one of the main limitations of repeat-pass
radar interferometry. The specific characteristics of different atmospheric situations
can be analyzed to identify which constituents form the main bottleneck. Both a
deterministic and a stochastic approach can be chosen to address this problem. In
the deterministic approach, the anomalous signal is derived from additional mete-
orological or radio-geodetic measurements, and subtracted from the interferometric
products. In the stochastic approach the unknown signal is modeled using a param-
eterization of its magnitudes and its spatial or temporal behavior.
4 Is it possible to analyze the atmospheric signal as a meteorological source of infor-
mation?
The problem of eliminating or modeling the atmospheric signal from interferograms
could possibly be inverted. Knowing the amount of deformation and the topographic
6 Chapter 1: Introduction

component in an interferogram implies that the resulting product (the differential

interferogram) reflects atmospheric delay variability only. From the ERS data set
alone, thousands of interferometric combinations are available for atmospheric stud-
ies, with as the main advantages the very high spatial resolution and phase accuracy.
Performing a meteorological interpretation of these interferograms could result in
improving the value of radar interferometry for a new discipline.

Research limitations

To narrow the scope, this study will focus on the error sources in monostatic space-
borne repeat-pass interferometry with main emphasis on differential interferometric
applications. Error analysis and interpretation of signal induced by the troposphere
is a key topic, combined with problems related to decorrelation and data processing.
It will not elaborate on errors induced by “phase unwrapping,” apart from discussing
the essential topics related to it. It will be assumed that the phase ambiguities can be
resolved without errors. Regarding SAR data processing, only the parameters that
directly influence the interferometric processing and the quality and interpretation
of the interferogram will be discussed.

1.4 Research methodology

In terms of methodology, different approaches are used to address the four main
research questions listed above. To parameterize the formation and quality of the
interferogram, the data acquisition and SAR focusing is regarded as a closed sys-
tem, from which the complex point statistics are given in terms of the probability
density function (PDF). Interferometric processing is described using Euclidean ge-
ometry and standard signal processing methods such as linear transfer functions and
spectral representations. A formulation of the relation between the interferometric
phase observables and different geophysical parameters is proposed using a linearized
Gauss-Markoff model to allow for future data adjustment and filtering.
Error propagation studies are based on Euclidean geometry for satellite orbit er-
rors, system simulations using ERS data characteristics and linear transfer func-
tions for the data processing. Various practical experiments have been performed,
e.g., for corner reflector movement and atmospheric signal delay. In these cases,
cross-correlations with triangulation, leveling, GPS, and various meteorological ob-
servations are performed.
A statistical interpretation of the effects of atmospheric refractivity variations is
performed using a series of interferograms created from 52 SAR acquisitions. Mod-
eling of the results is based on turbulence theory and power law behavior, which
is consequently translated to structure functions and covariance functions to derive
the variance-covariance matrix for atmospheric signal. Since the atmospheric signal
cannot be considered stationary for limited temporal and spatial intervals modeling
includes these non-stationary effects. Physical sensitivity analyses are performed
using forward modeling and ray-tracing experiments.
 1.5 Outline 7

1.5 Outline

The current chapter is an attempt to embed radar interferometry as a new geodetic

tool amidst the latest developments in geodesy. The main objectives addressed in
this book are formulated using a main problem statement, divided in four specific
research questions. Key methodologies are listed.
Chapter 2 is a review of radar fundamentals, extended to synthetic aperture radar
and radar interferometry. It serves as a general reference for data processing issues,
the use of supplementary data, and the physical aspects of signal propagation and
scattering. This chapter is not intended to completely cover the field, but to provide
proper reference needed for the subsequent chapters.
The chapters 3 and 4 present the description of radar interferometry in a standard
geodetic setting, using a Gauss-Markoff formulation. In such a mathematical model,
the relation between the observations and the unknown (geophysical) parameters is
described by the functional model, the main topic of chapter 3. This model is
then decomposed for three application areas: topographic mapping, deformation
mapping, and atmospheric mapping.
Chapter 4 comprises the stochastic part of the Gauss-Markoff model and covers
error and error propagation studies for the subsequent parts of data acquisition and
processing, as well as a unified model combining these different contributions. The
chapter starts with a short revisit of the necessary theoretical concepts, followed by
a single-point statistics part (section 4.2–4.5) and a multiple-point statistics part
(section 4.6–4.8). The latter part covers orbit errors and atmospheric errors.
Chapters 5 and 6 present the results of some case studies. Chapter 5 focuses on
deformation monitoring and presents the results of experiments involving corner re-
flectors, identification of coherent scatterers, subsidence monitoring, and earthquake
deformation monitoring.
Chapter 6 is self-contained and focuses on issues related to the use of radar interfer-
ometry for studies of mesoscale atmospheric dynamics. This is a novel application
of the technique, based on the quantitative analysis of lateral signal delay gradients
caused by atmospheric refractivity. Nevertheless, apart from these meteorological
applications, the geodetic analysis of the atmospheric signal can provide new insights
in the variability of refractivity, which influences all types of space-geodetic obser-
vations, such as GPS and VLBI. Especially the stochastic modeling of atmospheric
delay can benefit from these observations, as shown both in chapter 4 and 6.
Conclusions and recommendations for future research are given in chapter 7.

Hints for reading

This book is intended to be of some value for geodesists, solid earth geophysicists,
meteorologists, and radar scientists. For the first group of readers, it should bridge
the gap between the electro-technical background of radar, SAR, and radar interfer-
ometry and the basic “language” of geodesy. This involves (i) the translation to a
geodetic set of observation equations, (ii) the practical geodetic applications of the
8 Chapter 1: Introduction

technique, focused on accuracy and reliability, and (iii) the complementary value of
the technique amidst conventional geodetic techniques. For geodesists, chapters 3,
4, and 5 may be of direct interest. Chapter 2 discusses the necessary background,
whereas chapter 6 might be interesting to geodesists involved in space-geodetic tech-
niques where atmospheric propagation plays an important role.
For the second group, solid earth geophysicists, the applications and limitations of
the technique are important. After reading mainly chapters 4, 5, and the conclusions
of this book, they should be able to decide whether radar interferometry is a feasible
technique for a specific geophysical problem.
Meteorologists, the third group, may have technical as well as scientific motives to
read this work. Technical intentions are mainly based on the fact that the future will
make more and more data available, on a routine basis, in which the propagation
of radio signals from satellites can be used to infer properties of the atmosphere.
Chapters 3 and 4 discuss the basics between the atmospheric parameters and the
propagation of radio waves. The scientific interest of meteorologists in radar inter-
ferometric data could arise due to the unique characteristics of the data. The radar
data are not obscured by cloud cover, can be acquired during day and night, and
deliver quantitative information on the entire atmospheric column, in contrast to
many other meteorological imaging data. Another unique feature of the data is the
very high spatial resolution and the high accuracy of the integrated refractivity mea-
surements. In chapter 6, several case studies serve as examples for the atmospheric
value of radar interferometric data.
The last group, radar scientists, may refer to the references listed in chapter 2 to find
more technical aspects of radar data processing and interferometry. The evaluation
of the data accuracies and error sources, see chapters 3 and 4, especially regarding
the analysis of atmospheric signal, might be useful regarding future satellite mis-
sions. The design of future missions is influenced by the sensitivity of the radar for
disturbances and the surveying characteristics. For that matter, the required orbit
accuracies are of importance, as well as the sensitivity for tropospheric and iono-
spheric heterogeneities, the spatial range and magnitudes of these heterogeneities,
and the error propagation. This information is presented mainly in chapters 4 and 6.
Where numerical examples are discussed, radar parameters of the ERS SAR are
used, unless stated otherwise.
 Chapter 2
Radar system theory and interferometric
processing

This chapter reviews the basic concepts of radar and interferometry. It consists
of a short history of radar, SAR, and interferometry, followed by an instru-
ment description, listing the main components of a typical SAR instrument. A
section on image formation describes the main processing issues in the forma-
tion of a complex SAR image based on raw satellite radar data. A major part
of this chapter focuses on the interferometric processing cookbook, discussing
the main issues and problems. Differential interferometry is covered in the last
section of this chapter.
key words: Radar, Signal processing, Interferometry

2.1 Radar history and developments

“InSAR” is a nested acronym: Radio detection and ranging (radar), Synthetic Aper-
ture Radar (SAR), and interferometric SAR (InSAR). A short review of the devel-
opment of the subsequent techniques follows the acronyms in chronological order.

2.1.1 Radar

Radio detection and ranging (radar) refers to a technique as well as an instru-

ment. The radar instrument emits electromagnetic pulses in the radio and microwave
regime and detects the reflections of these pulses from objects in its line of sight.
The radar technique uses the two-way travel time of the pulse to determine the range
to the detected object and its backscatter intensity to infer physical quantities such
as size or surface roughness. A monostatic radar uses only one antenna, both for
transmitting and receiving, whereas in a so-called bistatic radar, the transmitting
and receiving antennas are physically separated (Skolnik, 1962)
Two landmark discoveries in the development of radar were Maxwell’s equations of
electromagnetism in 1873, in “Treatise on Electricity and Magnetism,” and Hertz’s
experiments in 1886. Hertz generated and detected the first known radio waves, and
discovered that these waves were subject to reflection or scattering (Buderi, 1996).
Although first radar systems were developed as early as 1903 for ship tracking and
10 Chapter 2: Radar system theory and interferometric processing

collision avoidance (the “telemobiloskop” of Hulsmeyer (1904)), pre-WW II military

considerations boosted the development of radar in the 1930s and 40s (Curlander
and McDonough, 1991).
The cavity magnetron was an important British invention which enabled the con-
struction of small radars with short pulses, high transmitter power, and cm-scale
wavelengths. Significant improvements of radar technology, triggered by the avail-
ability of the cavity magnetron, were achieved at MIT’s Radiation Laboratory during
the 40s, although similar developments occurred simultaneously at several locations
around the world (Buderi, 1996). Basically all current forms of radar can be traced
back to those early developments.
Although the technology mainly gained momentum through military applications
in those early years, civil and scientific applications arose quickly afterward. In
particular radio astronomy profited from the advances in technology. Earth-based
radars proved ideal for studying celestial bodies such as the Moon, Venus, Mars,
and the Sun. In January 1946, the first radar echo from the Moon was received
(Buderi, 1996). Using the radio-astronomy antenna’s in Arecibo and Goldstone,
radar observations of the planets Venus and Mars were acquired in 1961 and 1963
respectively (Goldstein and Gillmore, 1963; Goldstein and Carpenter, 1963), retriev-
ing four parameters of planetary bodies: velocity, time-delay, and intensity for the
normal and crossed polarization (Goldstein, 1964, 1969). Important results from
these measurements were the retrograde rotation of Venus and the improvement of
the astronomical unit.
The first tests to bring a radar into space started with JPL efforts in 1962, and lunar
radar observations from Apollo 17 proved successful in 1972. Nadir-looking radar al-
timeter measurements from earth orbiting satellites have been collected since Skylab
(1973) and GEOS-3 (1975). Venus was first visited by a radar on Pioneer 1, launched
1978. It carried a nadir-looking radar altimeter which enabled a global mapping with
an accuracy of about 300 m, albeit with a spatial resolution of about 50 km. The
soviet Venera 15 and 16 improved that resolution to about 2 km (Curlander and
McDonough, 1991; Henderson and Lewis, 1998).

2.1.2 Synthetic Aperture Radar

A specific class of radar systems are the imaging radars, such as Side-Looking (Air-
borne) Radar (SLR or SLAR) and later Synthetic Aperture Radar (SAR). The
side-looking geometry of a radar mounted on an aircraft or satellite provided range
sensitivity, while avoiding ambiguous reflections. The first SLAR’s were incoherent
radars—the phase information of the emitted and received waveforms was not re-
tained. The resolution in the flight direction was obtained by using a physically long
antenna, hence the name Real Aperture Radar (RAR). The practical restrictions on
the antenna length resulted in a very coarse resolution in the flight direction, degrad-
ing with higher flight altitudes, even though high-frequency (K-band: 0.75–2.40 cm)
radars were used, see eq. (2.3.4) on page 26.
Using a fixed antenna, illuminating a strip or swath parallel to the sensor’s ground-
track, resulted in the concept of strip mapping. Although modern phased-array
 2.1 Radar history and developments 11

antennas are able to perform even more sophisticated data collection strategies, e.g.,
ScanSAR, squint mode SAR, and spotlight SAR, the strip map mode is still the most
applied mode on current satellites.
The concept of using the frequency (phase) information in the radar signal’s along-
track spectrum to discriminate two scatterers within the antenna beam is generally
contributed to Carl Wiley in 1951 (Wiley, 1954; Curlander and McDonough, 1991).
Developments at the universities of Illinois and Michigan in the late 50s and early 60s
culminated in the concept of the synthetic aperture radar. The key factor for these
advances is the coherent radar, in which phase as well as amplitude are received
and preserved for later processing. Moreover, the phase behavior needs to be stable
within the period of sending and receiving the signal. As a result, an artificially
long antenna can be created synthetically using a moving antenna, combining the
information of many received pulse returns within the synthetic antenna length.
This methodology leads to a dramatic increase in azimuth resolution of about three
orders of magnitude.
Satellite remote sensing started with optical-mechanical scanners such as the Landsat
satellites, working in the visible and infrared parts of the electromagnetic spectrum.
Although these first spaceborne remote sensing images provided an unsurpassed
new perspective of the planet, some strong limitations became evident soon. Almost
continuous cloud cover, especially over tropical regions, lacking solar illumination of
the polar regions for half the year, and the new physical properties to be derived
from scattering in other parts of the spectrum triggered the development of satellite
SAR missions.
The experiences of mainly airborne SAR tests in the 60s and 70s culminated in
an L-band Synthetic Aperture Radar system onboard Seasat, a satellite launched
in June 1978 primarily for ocean studies. Although a short-circuit ended this mis-
sion prematurely, after 100 days of operation, the SAR imagery was spectacular,
and enabled SAR systems to be included in many satellites to come (Elachi et al.,
1982). Main highlights of Seasat were the wealth of geologic information and ocean
topography information that could be retrieved. The Seasat SAR imagery could be
processed optically as well as digitally (Henderson and Lewis, 1998).
Based on the success of Seasat, several Space Shuttle missions carried a synthetic
aperture radar, starting from 1981. The first instrument was the Shuttle Imaging
Radar (SIR) laboratory. SIR-A, an L-band instrument with an incidence angle of
50°, operated for 2.5 days in 1981 and provided valuable engineering experience.
SIR-B, an improved version of SIR-A, orbited the earth in 1984 and was able to
steer its antenna mechanically to enable different look angles. Whereas all data
from Seasat and SIR-A were recorded analogously on tapes, SIR-B was already
equipped with full digital recording.
Cosmos-1870 was the first (S-band) SAR satellite of the former Soviet Union, launched
in 1987 and orbiting at a height of 270 km. It operated for two years. ALMAZ-1,
the second satellite was launched in 1991, and operated for 1.5 years.
The first satellite SAR mission to another planet, Venus, was launched May 1989,
and named Magellan (Buderi, 1996). Magellan mapped 98% of Venus with a reso-
12 Chapter 2: Radar system theory and interferometric processing

lution of about 150 m from September 1990 until September 1992 (Henderson and
Lewis, 1998) (Curlander and McDonough, 1991), during three consecutive imaging
cycles. The combination of cycle I and III provided imaging from the same side,
with a 5°–25° difference in look angle. This geometry provided sufficient parallax to
obtain stereo-SAR elevation models with an accuracy of 150–200 m (Henderson and
Lewis, 1998). Due to the large orbit separation, interferometry was only possible at
a few locations were the orbits cross (Gabriel and Goldstein, 1988).
The European Space Agency (ESA) launched its first satellite equipped with a SAR
in July 1991: ERS-1 (European Remote Sensing Satellite). Although planned for
a 5-years life time, the satellite operated until March 2000. Designed primarily for
monitoring polar oceans and ice, it operated under a look angle of 20.3°, resulting
in strong topographic distortion. Nevertheless, the systematic data acquisitions, the
orbit control, and the data distribution policy resulted in impressive advances in
science, and major improvements in radar interferometry. An almost exact copy
of ERS-1, ERS-2, was launched in April 1995, which enabled a so-called “tandem-
mode” operation of the two satellites, a period of nine months in which ERS-2
followed ERS-1 in the same orbit, with a temporal spacing of 30 minutes. This
resulted in groundtracks which repeated with an exact 24 hour interval. After 1996,
ERS-1 acted as a backup satellite in hibernation mode, and was only activated
for special occasions, mostly related to SAR interferometry. Such occasions were
the eruption of volcanoes such as Vatnojöküll on Iceland, and severe earthquakes
such as the Izmit earthquake in 1999 (Barbieri et al., 1999). Both ERS satellites
operated from the same near-polar orbits at an altitude of 785 km. An image mode
is used to image the earth’s surface, while a wave/wind mode is used over oceans
(Attema, 1991). The satellite data are used for many environmental purposes (oil
slick monitoring, wind and wave field observations, transport of sediments, ice cover
and movement) over the oceans. Over land, the data reveal useful information
on land use, agriculture, deforestation, earthquake and volcanic deformation, and
general geophysics.
Japan started its spaceborne SAR program launching the Japanese Earth-Resources
satellite (JERS) in 1992. JERS was designed for solid earth remote sensing, and
carried an L-band (23.5 cm) radar with a look angle of 35°, more favorable regarding
topographic distortion. JERS operated until October 1998. The L-band acquisitions
proved to be well-suited for repeat-pass interferometric applications, due to their
reduced sensitivity for temporal changes in the scattering mechanisms at the earth’s
surface. Unfortunately, orbit control and maintenance was not as advanced as for
the two ERS satellites.
Radar observations from the Space Shuttle continued in 1994, with SIR-C/X-SAR,
a combined instrument developed by JPL, DLR, and ASI. The instrument orbited
the earth in Spring and Autumn, to detect seasonal differences in the images. The
mission operated for the first time in three frequency ranges: L-band, C-band, and
X-band, with 23.5, 6, and 3 cm wavelengths respectively. The combination of the
different frequencies extended and improved the interpretation of the data signifi-
cantly (Lanari et al., 1996; Melsheimer et al., 1996; Coltelli et al., 1996). Interesting
discoveries during the 1994 flights were the forgotten temples of Angkor, Cambodia,
 2.1 Radar history and developments 13

and the remains of an earlier version of the Chinese wall, close to the known one,
while not recognizable at the surface.
Started in 1995, the Canadian Space Agency (CSA) operates Radarsat, a SAR
satellite designed to perform especially Arctic observations of ice coverage, e.g., for
shipping route planning, apart from many other applications (Ahmed et al., 1990).
Radarsat has a ScanSAR mode and can acquire wide swaths. Although capable
of interferometric measurements, orbit control and maintenance is limited, which
restricts interferometric capabilities.
The experience with the repeat pass missions for topographic mapping, especially
the problem of temporal decorrelation and atmospheric disturbances, culminated in
the Shuttle Radar Topography Mission (SRTM) (Jordan et al., 1996). This Space
Shuttle mission was performed between 11 and 23 February 2000 and used a single-
pass configuration with a fixed 60 m boom to carry the two radar antennas. It
mapped all land masses between 60°N and 58°S using C-band, and tiles of this area
with a higher accuracy using X-band (Bamler et al., 1996b; Werner, 1998).

2.1.3 Interferometry

Compared with conventional geodetic techniques, one capability long remained out
of reach for radar: the measurement of angles. Similar to a single human eye, which
is essentially “blind” for the difference in distance to objects, it is impossible for a
radar or SAR to distinguish two objects at the same range—but different angles—to
the instrument. Nature readily provides the simple solution for the problem; the use
of two sensors. It worked with two eyes, why not use two radars?
This idea, and the use of the phase information, cleared the way for interferometry.
Using two SAR images, acquired either by two different antennas or using repeated
acquisitions, it is possible to obtain distance as well as angular measurements. The
use of the phase measurements (multiplicative interferometry) enabled the observa-
tion of relative distances as a fraction of the radar wavelength, and the difference
in the sensor locations enabled the observation of angular differences, necessary for
topographic mapping.
Basics of interferometry
Christiaan Huygens (1629-1695) developed the idea of the wave front in Traité de
la Lumière (Huygens, 1690). He observed that two intersecting light beams did not
bounce off each other as would be expected if they were composed of particles, and
that an expanding sphere of light behaves as if each point on the wave front were
a new source of radiation of the same frequency and phase. Interferometry of light
and other electromagnetic signals is based on the wave front concept.
Two distinct types of interferometry can be distinguished: additive interferometry
and multiplicative interferometry, see fig. 2.1. The signal resulting from the former
is obtained by the incoherent summing of the amplitudes of two input signals. This
results in so-called “amplitude” fringes, as in fig. 2.1 A. One problem with this type
of interferometer is that it is not possible to determine whether the interferometric
signal is increasing or decreasing. Another problem is that the accuracy is only a
14 Chapter 2: Radar system theory and interferometric processing

fraction of the amplitude cycle, which is often not very accurate. The application
of additive (incoherent) interferometry is usually a consequence of limitations in
detector technology, for example due the high frequencies of light. An example of
an additive interferometer is the famous Michelson-Morley interferometer (1886).1
In multiplicative interferometry (fig. 2. 1B) one is also able to measure the interfer-
ometric phase by coherent cross-multiplication of the two input signals. The phase
data imply a better accuracy, as the fraction of the phase cycle is much easier to
determine accurately. In 2D applications, a phase-cycle is often referred to as a
fringe.
Radar (SAR) Interferometry
Radio interferometry developed after the Second World War. In 1946, Ryle and
Vonberg constructed a radio analogue of the Michelson-Morley interferometer and
soon located a number of new cosmic radio sources (Tubbs, 1997). In the field of
planetary mapping, the range-Doppler configuration enabled the mapping of radar
reflectivity from iso-range lines and iso-Doppler lines (Evans and Hagfors, 1968).
What remained unknown was the ambiguity between reflections from the northern
and southern hemispheres of the planet. Rogers and Ingalls (1969) were the first to
use interferometry to resolve this ambiguity, using two antennas. For the moon, the
antenna beam width was small enough to resolve the north-south ambiguity, which
enabled Zisk (1972b) to use interferometry for measuring elevation differences (Zisk,
1972a; Shapiro et al., 1972). These applications of interferometry are still used today
(Margot et al., 1999a,b). For topographic mapping of the lunar poles, Margot et al.
(1999a) reported a horizontal resolution of 150 m and a height resolution of 50 m,
1
In 1868, the French physicist Fizeau (1819-1896) suggested using an interferometric method to
measure stellar diameters by placing a mask with two holes in front of a telescope’s aperture. He
argued that the fringes would vanish at a separation related to the size of the star. Fizeau’s ideas
were pursued unsuccessfully by Stephan and, using a different concept, successfully by Michelson.
 2.1 Radar history and developments 15

using two antennas at a distance of 20 km.

First experiments with airborne radar (SAR) interferometry for topographic map-
ping were performed by the U.S. military and a patent for obtaining elevation
from phase difference images was filed in 1971 (Richman, 1982; Henderson and
Lewis, 1998). Graham (1974) first published results with this method, using two
antennas and coherent additive interferometry with optical processing techniques.
Dual (single-pass) antenna airborne interferometry matured in the 80s, using digital
processing techniques, and using coherent multiplicative interferometry (Goldstein
et al., 1985). The first results, by Zebker and Goldstein (1986), produced interfer-
ograms, in which for every resolution element the amplitudes of the two images are
multiplied, and the phases are differenced. This resulted in a topographic map of
an area of 10 × 11 km, with an accuracy between 10 and 30 m. In 1987, Goldstein
and Zebker showed that using two antennas mounted in the flight direction of the
aircraft, along-track interferometry results in a sensitivity to motion of scatterers,
relative to the stationary background. The application of interferometry with satel-
lite data, using the repeat-pass method, where the satellite revisits an acquisition
area after a certain period, was first demonstrated in 1987 by Li and Goldstein, (Li
and Goldstein, 1987, 1990; Prati et al., 1990), using historic Seasat data, and by
Gabriel and Goldstein (1988) and Goldstein et al. (1988) using the shuttle’s SIR-B
data. The consecutive application of repeat-pass interferometry using aircraft data
was demonstrated by Gray and Farris-Manning (1993).
The estimation of topography had been the main focus for the early applications of
radar interferometry, yielding elevation accuracies comparable with optical methods,
albeit valuable due to the all-weather capability of radar. Nevertheless, the experi-
ments with the repeat-pass configuration clearly demonstrated an even more spec-
tacular application—deformation monitoring. The relative line-of-sight movement
of scatterers with respect to a reference location in the image could be measured as
a fraction of the wavelength, yielding cm to mm accuracies for L-band, C-band, and
X-band radars (Gabriel et al., 1989). A problem in this application is that—for an
effective baseline larger than zero—the deformation signal is always mixed with to-
pographic signal. A suitable solution to this problem was differential interferometry,
where the topographic signal obtained from a so-called topographic interferogram (or
a reference elevation model) was scaled to the baseline conditions of the deformation
interferogram and subtracted from it, yielding a differential interferogram (Gabriel
et al., 1989). The first demonstration of differential radar interferometry for map-
ping the displacement field of the Landers earthquake was reported by Massonnet
et al. (1993), who used a reference elevation model to remove the topographic phase
signal. Zebker et al. (1994b) reported the so-called three-pass method, studying
the same earthquake. This method uses two SAR acquisitions with a short tempo-
ral separation for retrieving the topographic phase signal, and the combination of
one of these two with a third acquisition. The latter interferogram, which spanned
the earthquake (the coseismic interferogram), was consequently corrected for to-
pographic contribution by the first interferogram. Massonnet et al. (1994b) showed
that even over time spans of one year, arid and undisturbed areas can still be used to
combine two SAR images interferometricly, although they argued that atmospheric
16 Chapter 2: Radar system theory and interferometric processing

delay signal might influence their results.

The only solid earth objective of ERS-1 was the detection and management of land
use. As interferometric applications of the SAR were not included in the design of the
satellite, the many studies on topographic and deformation mapping were received
with much interest by the scientific community, see (Hartl, 1991; Prati et al., 1993;
Hartl et al., 1993; Hartl and Thiel, 1993; Prati and Rocca, 1993; Prati et al., 1993;
Prati and Rocca, 1993; Small et al., 1993; Werner et al., 1993; Prati and Rocca,
1994a). The following section gives an overview of satellite SAR missions and their
interferometric applications.

2.1.4 Missions and applications

Table 2.1 lists the most important SAR missions, showing their main characteristics
in terms of life time, orbit, and SAR performance. The satellites indicated by a star
are capable of interferometric applications. The dates of operation give an estimate
of the historical archive of SAR data, and the orbital repeat-period indicates the
sensitivity to temporal decorrelation and deformation rates. The altitude can be
used as an indication for orbit stability, cycle period, and the influence of drag
effects.
The SAR frequency and bandwidth are used to determine the radar wavelength, and
hence its sensitivity to surface displacement, topographic height, temporal decorre-
lation, and range resolution, using equations discussed in sections 2.3 and 2.4. Using
the incidence angle, height, frequency, and bandwidth parameters, the critical base-
line can be calculated, which expresses the maximum horizontal separation
of the two satellite orbits in order to perform interferometry, see eq. (4.4.11) on
page 102. Finally, the swath width fixes the feasible spatial scales of the phenomena
of interest.
The geodetic applications of spaceborne repeat-pass SAR interferometry can be cat-
egorized in roughly four disciplines (i) topographic mapping with a relative accuracy
of 10–50 m, (ii) deformation mapping with mm–cm accuracy, (iii) thematic map-
ping based on change detection, and (iv) atmospheric delay mapping with mm–cm
accuracy in terms of the excess path length.
Topographic mapping
Applications of spaceborne radar interferometry for creating digital elevation models
(DEMs) have matured from early experiments, see Zebker and Goldstein (1986), to
a standard geodetic tool in one decade. Today, commercially available software can
be used by non-expert users, and companies provide services in creating geocoded
DEMs from the available SAR data. Especially data from the ERS-1/2 “tandem
mission” are currently used for these purposes (van der Kooij, 1999a). Due to
the regular acquisitions of the ERS satellites, aimed at full coverage of the world,
mosaicing of many interferometric elevation models enables the mapping of large
areas, up to continental scales. Main applications of the elevation models include
telecommunication, hydrological mapping and flooding predictions, cartography, and
geophysics. The availability of topographic maps (status 1997) is shown in fig. 2.2,
 2.1 Radar history and developments 17

stressing the lack of detailed topographic maps in many parts of the world.
Radar interferometric measurements of topography may seem similar to optical
(stereographic) methods at first glance. Both methods need two “images” to infer
height differences. The main technical difference, however, is that optical techniques
are based on parallax, which is directly based on angle-measurements, whereas radar
is a “ranging” device, which measures distances. Parallax measurements require the
identification of homologous features in the terrain, which poses problems in, e.g.,
snow areas in polar regions or sand areas in deserts. Moreover, optical techniques
need illumination by the sun and cannot penetrate through cloud cover. For ex-
ample, in Europe only 10% of all acquired optical data are useful for topographic
mapping due to cloud cover, whereas high-latitude regions are not illuminated by
the sun for several months per year.
18 Chapter 2: Radar system theory and interferometric processing

The main problems with topographic mapping using repeat-pass interferometry are
layover, foreshortening, shadow, surface decorrelation (especially due to vegetation
and the decrease of coherence over the time interval between the acquisitions), and
the atmospheric signal in the data. The effective baseline length influences the
latter two error sources. For example, a rather long baseline increases noise in the
data due to geometric decorrelation, whereas it reduces the effect of atmospheric
disturbances in the derived DEM dramatically, and vice versa. Note that single-pass
interferometry is not influenced by these error sources, since the data are acquired at
the same time, and atmospheric signal will cancel in the interferometric combination.
The SRTM mission, using single-pass interferometry, is therefore specially designed
to measure topography, yielding elevation data that are better and more reliable
than repeat-pass data, albeit over a limited part of the world (Jordan et al., 1996;
Bamler et al., 1996b).
Deformation monitoring
Since the observation of ground motion over agricultural fields by Gabriel et al.
(1989), many applications of (differential) InSAR for deformation monitoring have
been developed. As the scaling effect of the baseline, which influences the sensitivity
for topographic mapping, is not present in deformation mapping, accuracies can
reach the sub-cm level.
Applications extend to many forms of surface deformation. Table 2.2 shows that the
applications can be organized in several categories. Seismic applications include the
investigation of earthquakes, faults, and tectonics. The use of two acquistions to form
an interferogram results in the capability of measuring preseismic, coseismic, and
postseismic deformation. These terms are defined as deformation preceding a seismic
 2.1 Radar history and developments 19

event (two images before), spanning a seismic event (one image before and one after),
or following a seismic event (two images after), respectively. These possibilities, using
archived satellite data, enable geophysical and hazard studies with an accuracy and
resolution unprecedented by any other geodetic technique. Coseismic deformation on
scales of cm-m are observed by, e.g., Massonnet et al. (1993); Zebker et al. (1994b);
Peltzer and Rosen (1995), Massonnet and Feigl (1995a,b), and Peltzer et al. (1999).
Especially the combination of the radar data with GPS observations extends the
analysis and interpretation of earthquakes considerably (Segall and Davis, 1997).
Postseismic deformation has been observed by, e.g., Massonnet et al. (1994b, 1996b)
and Peltzer et al. (1996).
In terms of volcano deformation monitoring, the same three categoeries can be dis-
tinguished. In case of volcanic eruption, pre-, co-. and post-eruptive measurements
can be performed. Of course, the availability of an archive of pre-eruptive SAR
observations for many volcanoes is important in the development of hazard fore-
casting capabilities. Although the detection and monitoring of deforming volcanoes
is of major importance, it might be equally important to systematically monitor
“sleeping” volcanoes that could pose a potential threat (Amelung et al., 1999).
Land subsidence, for example caused by mining activities, is a third category con-
cerning deformation measurements. Many areas in the world are affected by sub-
sidence (or uplift) due to the extraction of water, gas, oil, salt, or other mineral
resources. Moreover, construction works such as tunnels often cause subsidence re-
sulting in damage to the infrastructure. The feasibility of radar interferometry for
these applications depends on the subsidence rates, i.e., deformation gradient, in
combination with the influence of decorrelation due to land use or vegetation and
of atmospheric signal in the interferograms. Although the temporal decorrelation
poses a significant limitation for obtaining continuous deformation maps, it is ob-
served that urban areas or specific natural or anthropogenic features remain coherent
for extended time intervals, (Usai, 1997; Usai and Hanssen, 1997; Hanssen and Usai,
1997; Usai and Klees, 1998; Ferretti et al., 1998; Usai and Klees, 1999a,b). Recent
advances using stacks of coregistered SAR images show that many permanent (co-
herent) scatterers can be identified, especially over urbanized areas (Ferretti et al.,
1999b, 2000). Since conventional coherence estimators use estimation windows that
are often too wide to identify a single scatterer, the systematic analysis of many
images is inevitable for a robust identification procedure.
Glacier and ice motion in the order of meters can be conveniently studied using
interferometry as well, see e.g. Goldstein et al. (1993). Due to the inaccessibility
of the terrain, e.g., Greenland or the Antarctic, the interferometric measurements
provide invaluable information on glacier dynamics, which can be used as input for
more advanced problems such as global warming and sea-level rise. Depending of
the speed of the ice sheet and glacier velocity, a 35 day repeat interval may be too
long for the SAR images to stay within the grid distance needed for coregistration
conditions. In such cases, additional SAR techniques such as speckle tracking (Gray
et al., 1999) may be neccesary in combination with the interferometric combination
of small patches in the SAR data. Many studies in this field have been performed,
see table 2.2.
20 Chapter 2: Radar system theory and interferometric processing

Continued on page 21
 2.1 Radar history and developments 21

Other, more specific, applications of deformation monitoring include the amount

of mass extraction from a quarry, see Hartl and Thiel (1993) or the movement of
man-made corner reflectors, see Hartl and Xia (1993), Hartl et al. (1993), and Prati
et al. (1993).
Thematic mapping
Changes in the scattering characteristics of the ground surface between the two radar
acquisitions can result in a change in the coherence—the degree of similarity between
the two images—in the interferogram. Whenever the scattering mechanisms remain
unchanged, the degree of coherence is highest, and only influenced by the acquisition
geometry, data processing, and thermal noise.
The analysis of the decrease in coherence associated with different types of land
use has been reported by Zebker and Villasenor (1992); Askne and Hagberg (1993);
Hartl and Xia (1993); Wegmüler et al. (1995); Yocky and Johnson (1998). Spe-
cific applications include: forestry, (Wegmüller et al., 1995; Wegmüller and Werner,
1995; Wegmüller and Werner, 1997), flood monitoring, (Geudtner et al., 1996), lava
streams, (Rosen et al., 1996; Zebker et al., 1996; Lu and Freymueller, 1999), ice
penetration, (Hoen and Zebker, 2000), land slides, (Rott et al., 1999), and fires, soil
moisture changes, or vegetation growth.
A problem in the analysis of the interferometric coherence measurement is the sep-
aration of the different sources of decorrelation. For example, it is necessary to
separate geometric decorrelation due to baseline characteristics from temporal decor-
relation in order to make a correct interpretation of the physical phenomena (Hoen
22 Chapter 2: Radar system theory and interferometric processing

and Zebker, 2000). Zebker and Villasenor (1992) and Gatelli et al. (1994) showed
that for surface geometric decorrelation, a linear model can be used, relating the
perpendicular baseline to the amount of decorrelation. Whenever volume decorre-
lation occurs, e.g.. in forest or ice applications, this model needs to be extended,
see Hoen and Zebker (2000).
Atmospheric delay mapping
The sensitivity of the radar signal delay for atmospheric refractivity variation is
usually considered to be a nuisance in radar interferometry. However, the abundance
of SAR acquisitions over areas where topography is known, and surface deformation
is absent, makes the interpretation of atmospheric signal in the data an interesting
goal for meteorology and the study of atmospheric dynamics. Nevertheless, due to
the low repetition frequency of current SAR satellites, these applications will have
a rather opportunistic character. Current studies therefore focus on demonstration
and validation of the technique, referred to as Interferometric Radar Meteorology
(IRM), rather than providing operational capabilities. Some demonstration and
validation studies are discussed in chapter 6 (Hanssen et al., 1999, 2001, 2000b).

2.2 Sensor characteristics

Synthetic aperture radars are available in many different configurations, although the
basic modules are usually the same. Most important, to enable SAR (and InSAR)
processing, the radar needs to be coherent, at least within the time span between
sending and receiving one pulse. In a coherent radar the phase of the transmitted
signal is retained and used as a reference against which the returned signal can be
compared. Here we will discuss essential hardware components of the instrument,
using the Active Microwave Instrument (AMI) on board of the ERS satellites as an
example.
The SAR is one of the two functions of the AMI, which consist of a 10 × 1m
area antenna, a SAR processor, a pulse generator, transmitter, and some other sub-
systems. The second function of the AMI, the wind scatterometer, uses several
common sub-systems. As a result, both functions cannot be operated in parallel.
Due to the power consumption of the SAR in image mode, it can only be operated
12 minutes per 100 minute orbit, of which maximally 4 minutes in eclipse (Attema,
1991).
During its operation, the SAR processor initiates the command to generate a radar
pulse with a duration of This pulse is generated and since it needs to
be shaped into a “chirped pulse” with varying frequency, the signal is fed into a
dispersive delay line. This device produces a linear frequency modulated (FM)
chirp, over a range of 15.5 MHz. The rate at which consecutive chirps are generated
is the pulse repetition frequency (PRF), which is programmable in the range 1640–
1720 Hz. The typical value is 1680 Hz, resulting in 10 transmitted pulses before
the first echo is received, see fig. 2.5. The chirp is passed to the transmitter and
up-converter part, which shifts the signal to a carrier frequency of 5.3 GHz. The
high-power amplifier amplifies the signal by approximately 45 dB, until it reaches
 2.3 Image formation 23

the desired peak power of 4 kW for the phased-array antenna. It then passes via
the waveguide to the antenna subsystem, consisting of a circular assembly and the
SAR antenna. The circular or switch assembly assures that the high-power transmit
signal is isolated from the sensitive low-noise amplifier for the echo signal.
When the echo signal is received by the SAR antenna, it is routed via the waveguide,
through the circulator assembly, to the receiver. The received radar echo is then
amplified in a low-noise amplifier and mixed with the local oscillator signal to provide
a down-converted signal at the intermediate frequency of 123 MHz. Apart from the
echo, also a sample of the transmitted waveform is down-coverted, which serves as
the replica signal for on-ground SAR processing. After routing these signals through
a pulse compressor they are sent to the analog-to-digital I/Q converters, producing
the real (in-phase) and imaginary (quadrature) part of the signal. Although the
SAR detector uses 6 bits for the replica and calibration pulses, it uses 5 bits for the
image data, i.e. 5 bits (I), 5 bits (Q). With a sampling frequency of 18.96 MHz,
this produces 10 bits of digital code every 52.74 ns, which are directly downlinked to
receiving stations on earth. Usually, these values are stored in one byte for standard
processing, i.e., 1 byte (I) and 1 byte (Q).
The Doppler effect due to the satellites velocity in an inertial reference system, in
combination with the earth’s rotation in the same system, introduces a frequency
shift in the received data. Since the beam width is small, these frequency shifts are
limited to a range of about 1500 Hz, well within the PRF of 1680 Hz. However,
since the additional effect of the earth rotation varies with latitude, the whole band-
width varies around the zero-Doppler location. This effect is reduced by applying
yaw-steering, where the total antenna beam is electronically shifted forward and
backward, in a range from 0° at the poles to about ±4° at the equator (Attema,
1991; Alaska SAR Facility, 1997). This results in a Doppler Centroid frequency that
is less than 900 Hz from the zero-Doppler direction (Bamler and Schättler, 1993).

2.3 Image formation

The antenna of the ERS-SAR is aligned along the satellite’s flight path to direct a
narrow beam sideways and downwards (20.3° to the right of nadir) onto the earth’s
surface to obtain strips of imagery of about 100 km in width, see fig. 2.3. Imagery
is built up from the time delay and strength of the return signals, which depend
primarily on the roughness and dielectric properties of the surface and its range to
the satellite.
The “synthetic” aperture of SAR uses the redundancy in subsequent pulse returns
that elucidate a single point scatterer. The inverse problem of reconstructing the
scatterers response from a series of pulse return signals is referred to as SAR focusing
or synthetic aperture processing. This methodology, first demonstrated by Graham
(1974) based on the concepts of Wiley (1954), improves azimuth resolution from the
4.5 km beam width for a single pulse to approximately 5 m for the full synthetic
aperture.
First focusing has been developed using optical processors, based on the concepts
of holography (Hovanessian, 1980). Currently, all processors are electronic (digital).
24 Chapter 2: Radar system theory and interferometric processing

Several electronic algorithms for SAR focusing have been developed: range-Doppler
(Bennett and Gumming, 1979; Wu et al., 1981; Curlander and McDonough, 1991),
seismic migration (Prati et al., 1990), the PRISME architecture (Massonnet et al.,
1994a), and chirp scaling (Runge and Bamler, 1992; Raney et al., 1994). See also
references in Bamler (1992), Massonnet and Feigl (1998), and Otten (1998). The
ERS SAR data used in this study were either preprocessed to SLC-format by one of
the ESA PAFs, or processed using the SAR processor developed at JPL and Stanford
 2.3 Image formation 25

University, which applies a range-Doppler algorithm (Zebker et al., 1994b).

The azimuth resolution of a focused SAR image is closely related to the dimensions
of the radar antenna. A small antenna with a broad beamwidth contains high-
frequency information of a point scatterer’s response, resulting in a broader azimuth
bandwidth (Bamler and Hartl, 1998). The bandwidth is in the order of
resulting in an azimuth resolution of about Although it might seem that
reducing the antenna size might increase resolution to infinity, the effective gain of
the antenna is proportional to the square of its aperture. Therefore, a reduction
in antenna size will be a trade-off with the maximum attainable signal-to-noise
ratio (Henderson and Lewis, 1998).
In the following subsections, the process of image formation or focusing is briefly
reviewed. As mentioned above, there are many SAR processing algorithms, and
some of those regard SAR image focusing directly as a two-dimensional problem.
Nevertheless, here we try to describe the process as consisting of smaller steps. We
start with describing the antenna pattern, then describe the range pulse waveform,
and finally the range compression. Then the Doppler centroid concept is introduced,
followed by range migration correction and azimuth focusing. This section concludes
with the imaging geometry of a single SAR image.

2.3.1 Antenna pattern

A rectangular antenna with length and width (see fig. 2.3) and a uniform
current density produces a (normalized) antenna pattern as a function of the off-
center beam angle see Olmsted (1993); Curlander and McDonough (1991): Zebker
(1996):

where and are the off-center angles in range and azimuth direction respectively.
Evaluating eq. (2.3.1) in azimuth direction, for constant this pattern is shown in
fig. 2.4.
Usually, in radar systems the energy pattern is considered constant between the half
power (3 dB) angles. In the pattern in range, the half power is reached at
and The beam w i d t h s a n d are therefore

For ERS-1/2, using the characteristics from Table 4.2, page 101, the theoretical
values for the beamwidths are

In fact, for ERS-1/2 the beamwidth of the main lobe is slightly broadened in range
direction in order to get the power distributed more evenly across the full swath.
26 Chapter 2: Radar system theory and interferometric processing

This results in practical beam widths of 5.4° and 0.228° in range and azimuth,
respectively (Attema, 1991).
The radar footprint width in azimuth direction at a distance R is now

i.e., approximately 4.8 km for ERS. For focused spot scanning, or real-aperture
radar (RAR), this is the maximum attainable resolution (Fitch, 1988). In range,
the footprint width or swath width is determined by the time interval of the pulse
return registration.

2.3.2 Range modulation

The transmit/receive configuration of the ERS-SAR is sketched in fig. 2.5. The radar
emits short, high-energy microwave pulses to earth and records the echoes from each
pulse. A monostatic radar uses one single antenna to transmit as well as receive the
pulses. The pulse rate is defined by the pulse repetition frequency (PRF=1680 Hz).
The pulse waveform can be written as (Bamler and Schättler, 1993)

where is the complex envelope and is the carrier frequency (5.3 GHz). The
resolution in range i.e., the shortest range difference at which two scatterers can
 2.3 Image formation 27

be distinguished, for a rectangular envelope depends on the pulse length

and therefore a shorter pulse envelope will result in a higher resolution in range
direction. On the other hand, to obtain a high-SNR radar image, high peak powers
are desirable, since the energy of the received pulse is orders of magnitude
smaller than the energy of the emitted pulse. These two considerations, short pulses
vs. high peak powers, are conflicting since the amount of energy the instrument can
emit in a finite time span is limited, and as a consequence there is a limit in reducing
the length of the pulse. For ERS, with we find the minimal distance
between two resolvable points in range direction to be or 14 km in
ground range. Note that this is independent of the altitude of the satellite.
The problem of the limited pulse length is circumvented by phase-coding the enve-
lope. In a linear frequency modulated (FM) chirp waveform the frequency of
the pulse increases linearly with slope i.e.,

Deriving the phase signal from (2.3.7) by integration yields or

omitting the integration constant. The bandwidth ranging from

is now defined as It can be shown that using matched filtering on the
received pulse, also referred to as a pulse or chirp compression technique, the effective
range resolution improves to

corresponding to ~25 m in ground range (Curlander and McDonough, 1991). See

Papoulis (1991) for the matched filter principle. The range enhancement factor can
be found from the time-bandwidth product:

which is 576.6 for an ERS SAR image. This resolution corresponds with an effective
compressed pulse length
The radar echo can be described as the convolution of the transmitted waveform and
the surface reflectivity and corresponds with one line in range direction. Considering
the echo of a point scatterer with phase shift the sensor receives a delayed replica
of the transmitted waveform:

where is the round-trip time between transmitting and receiving. Amplitude

scaling factors are omitted here. The linear operation of quadrature demodulation,
28 Chapter 2: Radar system theory and interferometric processing

using the local oscillator, shifts the signal to a frequency band centered at zero-
frequency, which results in a strict linearity in further signal processing (Curlander
and McDonough, 1991, p. 136). This yields

On board of the satellite, the signal is digitized to complex samples and downlinked
to a receiving station, along with a host of other engineering data (chirp replica,
noise measurement, calibration pulse) (Alaska SAR Facility, 1997).

2.3.3 Range or chirp compression

The first step in the SAR processing involves the improvement of the resolution in
range direction, by compressing the chirped wave form discussed in section 2.3.2.
The matched filtering procedure requires the replica of the transmitted chirp. These
replicas can (i) be computed, the so-called range reference function (RRF), or (ii)
directly retrieved from the downlinked data, as discussed in section 2.2. In the
raw ERS data, a replica of the transmitted chirp is available every 24 pulses. To
compute the RRF, eq. (2.3.8) is evaluated, using points, where is the sampling
frequency and is the pulse length. See fig. 2.6 for the shape of the reference
 2.3 Image formation 29

function.
The chirp or RRF is usually weighted with a Hamming filter to reduce sidelobe
effects, zero-padded to the next higher power of 2, and transformed to the Fourier
domain. The zero-padded raw SAR data of one pulse are Fourier transformed as well,
and multiplied with the RRF. Transforming this product back to the time domain
yields the range compressed signal (Curlander and McDonough, 1991; Zebker, 1996;
Price, 1999).

2.3.4 Doppler centroid

The antenna beamwidth in azimuth, causes a point scatterer to be imaged within

the beam for a number of consecutive pulses. During these pulses its relative velocity
with respect to the radar changes, causing a Doppler-like effect. Strictly, the analogy
with the Doppler effect is incorrect, as this is a continuous phenomenon, whereas
SAR acquisitions can be regarded as a discrete stop-and-go process (Bamler and
Schättler, 1993).
The antenna pattern of the ERS-SAR is aimed almost exactly perpendicular to the
sum of the flight direction vector and the velocity vector due to the earth’s rotation.
Nevertheless there will always be a slight squint angle of the spacecraft, steering
the antenna pattern away from the zero-Doppler direction, cf. fig. 2.7.
The Doppler frequency for a target approaching with velocity is In
the SAR geometry in fig 2.7A and B this is
30 Chapter 2: Radar system theory and interferometric processing

where is the velocity vector of the spacecraft in an earth-fixed frame and is the
look angle. Note that this equation determines the sign of the phase observations
in the sequel, since a velocity towards the sensor, or a decrease in range results
in a positive Therefore, Since the frequency
is obtained by differentiating the phase, with we find that a
decrease in phase corresponds with an increase in range, i.e.,
The center frequency of the passage of a point scatterer through the antenna beam is
termed the Doppler centroid frequency, whereas zero-Doppler, denotes
the direction in which the Doppler frequency is equal to zero. This direction is
perpendicular to the flight direction. Often the phase data in the focused SAR
image are deskewed, implying that the phase values correspond with the zero-Doppler
phase. Deskewed, or “zero-Doppler” processing means that the data are always
observed effectively perpendicular to the flight track, which is convenient for keeping
track of time, during geocoding, and coarse coregistration, although these steps can
be applied during the post-processing as well.
The variation of the Doppler frequency during the passage of the scatterer through
the beam is expressed by the Doppler bandwidth, For a narrow antenna beam
in azimuth direction, the Doppler bandwidth is well approximated by (Raney, 1998)

where is the azimuth beamwidth of the antenna and is the relative spacecraft
velocity. The radar PRF must be sufficient to sample the Doppler spectrum unam-
biguously, so that The Doppler bandwidth determines the azimuth
 2.3 Image formation 31

resolution of the focused SAR image, using

A typical Doppler bandwidth for ERS is 1377 Hz, depending on the definition of
the beamwidth. Note that higher azimuth resolutions are possible using a shorter
wavelength, a wider beamwidth, or a lower (faster) orbiting spacecraft.
In fig. 2.7B, a surface element in shaded area 2, after range migration, will have
a slightly altered phase value for the next pulse. In that time, the imaged area
will be almost the same, the only difference being the relative velocity towards
the sensor. This way, we can use the phase difference between neighboring pulses,
for the same range distance to estimate the Doppler centroid coarsely. Averaging
a number of those differences will give a reasonable estimate, as demonstrated in
fig. 2.8. More accurate algorithms are discussed in Madsen (1989). In the example
of fig. 2.8, an average phase shift of –1.83 radians is observed from the data, which
corresponds with –0.29 cycles. With a PRF of 1679.902 Hz, this corresponds with
an estimated Doppler Centroid Frequency of –488.92 Hz. The relative velocity of
the spacecraft is 7554.27 m/s, and using eq. (2.3.13) yields a squint angle
degrees. For focusing the SAR data, is important for SNR maximization and
for suppressing azimuth ambiguities (Geudtner, 1995). For the focused SAR data,
is equivalent to a spectrum that is shifted in azimuth direction. Such shifts
are important to consider for interferometric applications, e.g. during interpolation
procedures.
Note that varies over range, as sketched in fig. 2.7C. This variation is caused by
the variation of the incidence angle over range, cf. eq. (2.3.13), in combination with
complicating factors such as earth rotation and topography. Scatterers that have
a velocity component in the direction of the radar appear displaced in the focused
image.
32 Chapter 2: Radar system theory and interferometric processing

2.3.5 Range migration

Compare the passage of a scatterer through the beamwidth, area 2 in fig. 2.7B, with
the position of scatterers in a range bin, area 1 in fig. 2.7B. It is obvious that the
single scatterer appears shifted in different range bins in consecutive pulses, even if
the data are acquired under zero-squint. This effect (range migration) makes SAR
focusing an inherent 2D process. Although different concepts are available for solv-
ing this problem, a common solution is to shift the returns from range cells to a
common range bin in order to apply 1D transforms. This procedure is evaluated in
blocks of range-compressed data, Fourier transformed in azimuth. The range migra-
tion correction results in a non-integer shift of the data, requiring sine-interpolation
(Curlander and McDonough, 1991). During range migration correction it is nec-
essary to decide between zero-Doppler (deskewed) processing or Doppler centroid
(skewed) processing, as both strategies result in different migration corrections.

2.3.6 Azimuth focusing

Azimuth compression or focusing is a procedure analogous to range compression.

The concept of matched filtering, the convolution of a reference function to the
received signal can be applied using the chirp as reference function in range and
the predicted Doppler-rate (or phase-history) of a scatterer as reference function in
azimuth. The azimuth reference function is computed for each range bin and Fourier
transformed. The product of the transformed filter with the Fourier transformed raw
data is transformed back to the time domain, yielding a focused SAR image. The
basic concepts of this methodology can be found, e.g., in Curlander and McDonough
(1991).
In effect, focusing the SAR data yields a final azimuth resolution which is dependent
of the azimuth time bandwidth and the velocity of the sensor, see eq. (2.3.15).

2.3.7 Multilook processing

Multilook processing during image focusing can be applied to improve radiometric

accuracy at the cost of image resolution. This is usually applied in the range-Doppler
domain, by segmenting the Doppler spectrum in a number of subsets, or looks, and
summing these (independent) looks, see Elachi (1988); Curlander and McDonough
(1991). For interferometric applications single-look processing is commonly applied,
using the full azimuth bandwidth. In the interferometric SAR processing, the term
multilooking is also used for simply averaging a number of resolution cells to improve
phase statistics, see section 2.5.6. Phase statistics are discussed in detail in chapter 4.

2.3.8 Imaging geometry

The reference frame of a focused SAR image is spanned by the range-azimuth co-
ordinates. Since the radar is side-looking, terrain elevation will result in geometric
distortions in the SAR image. In fact, even the variation in the projection of the ref-
erence surface (ellipsoid) in range direction causes geometric distortions, due to the
varying incidence angle. Figure 2.9 shows the effects of foreshortening (A), layover
 2.3 Image formation 33

(B), and shadow (C), caused by the oblique viewing geometry. The radar coordinate
system can be transformed to a ground-range/azimuth coordinate system situated
on the ellipsoid.
Resolution and posting
Resolution is defined as the minimal distance at which two distinct scatterers with
the same brightness can be uniquely discerned as separate signals (Born et al., 1959).
For any system, the fundamental resolution is equal to one over the system band-
width, which corresponds with the width of the impulse response. These measures
of time can be easily transformed into spatial measures. In range this is expressed in
eq. (2.3.9), in azimuth in eq. (2.3.15). A resolution cell is defined as the illuminated
area responsible for the radar reflection data mapped to a single pixel. Therefore,
the resolution cell size is determined by the resolution in range and azimuth.
Pixel size, on the contrary, is a confusing measure, since it is often used in the same
context as resolution. A pixel, however, is an infinitesimally small point in which
34 Chapter 2: Radar system theory and interferometric processing

the complex digitized signal value corresponding with a resolution cell is centered.
A pixel, by definition, does not have a physical size, although the representation of
a grid of pixel values as an image might suggest so. An acceptable alternative for
pixel size is posting, the spatial distance between two pixel-nodes, or pixel spacing.
For systems that have the same bandwidth as sampling frequency, posting and res-
olution cell size have identical values. For most SARs, however, signal processing
algorithms require oversampling, or correlation between adjacent pixel values. In the
spectral domain this corresponds with a system bandwidth that is smaller than the
Nyquist frequency. For ERS focused SAR images, an oversampling ratio of ~1.223
is used, corresponding with 50% overlay between adjacent resolution cells.

2.4 SAR interferometry

Comparable to a single static radar, the range direction information in a synthetic

aperture radar only conveys information on the distance to the sensor at a certain
azimuth time. The distance measurement follows from the time observations of
the pulse returns by the local oscillator, resulting in a relative accuracy of half the
posting in range direction, typically 4–5 meters. As the phase information of a
resolution cell has a uniform probability density function, see chapter 4, it does not
contain any useful information.
Figure 2.10A shows that distance measurements are not capable of distinguishing
 2.4 SAR interferometry 35

two points P' and P at the same slant range, but displaced horizontally in ground-
range. Although the backscatter intensity can give an indication for the presence of
topography, this information cannot be made quantitative to a high accuracy. It is
evident that this problem could be solved by the observation of angular differences
between a point P' at a reference body and a point P at a height above this
reference body, with the same range to the sensor. This is equivalent with mea-
suring cumulative angular differences between neighboring resolution cells. In fact,
this is what SAR interferometry provides by observing both points from a slightly
different geometry, see fig. 2.10B. The effective distance between the two sensors,
measured perpendicular to the look direction, is referred to as the perpendicular or
effective baseline, Because the instrument is not capable of directly measuring
the small angular differences, this information needs to be derived from the distance
measurements between both sensors and the resolution cell on earth, applying some
simple trigonometry as indicated in fig. 2.10B. Thus, the basic problem of SAR
interferometry is the determination of these distance differences.
The required accuracy for measuring the distance differences between sensors and
resolution cell is in the mm-range. As the ranging information determined by the
range resolution is three orders of magnitude worse, it is not applicable for this
type of accurate observations. The phase observations of the received echo provide
the solution for this problem. Disregarding atmospheric propagation delay for the
moment, the phase observation for a single resolution cell can be regarded as the
sum of two contributions: the phase proportional to the distance and the phase
due to the scattering characteristics of the resolution cell. Although the scattering
phase component is unpredictable, it is a deterministic quantity, i.e., if the phase
measurement would be repeated under exactly the same conditions, it would yield
the same result. Under these circumstances we state that the imaging is coherent.
The degree of coherence is a direct measure for the similarity between the two
observations. As a consequence, the phase difference between two sensors for a
coherent system is only dependent on the difference in range, as the scattering phase
contributions cancel.
In practice this method imposes high demands on the geometric configuration—
parameterized by the spatial baseline—and the interferometric processing of the
SAR data. Moreover, for repeat-pass interferometry the temporal separation of
the two acquisitions—referred to as the temporal baseline—can result in changing
scattering characteristics due to weathering, vegetation, or anthropogenic activity.
All these parameters affect the degree of coherence up to total decorrelation.
In the following sections we will briefly discuss the influence of the parameters men-
tioned above. A more thorough mathematical treatment of these concepts can be
found in chapters 3 and 4.

2.4.1 Phase observation, topographic height, and surface deformation

Using the interferometric configuration sketched in fig. 2.10B, we derive the physical
and the geometrical relationships between the two phase observations to obtain
topographic height and surface deformation estimates.
36 Chapter 2: Radar system theory and interferometric processing

Both SAR images are composed of a regular grid with complex values, or phasors,
and which can be decomposed in an amplitude and a phase component using

After aligning and resampling the grid to corresponding locations in the grid,
complex multiplication yields the complex interferogram:

The observed phase values and in the two images for resolution cell P are

where and are the geometric distances and and are the con-
tributions of the scattering phases in both images. For now, we ignore phase con-
tributions due to signal propagation delay. The origin of the minus sign has been
discussed in section 2.3.4. In case the scattering characteristics are equal during
both acquisition, i.e., the interferometric phase can be written as

and its derivative is

Geometrically, the path length difference can be approximated as (cf. fig. 2.10B)

This approximation is known as the far-field or parallel-ray approximation, see Ze-

bker and Goldstein (1986). Due to the phase ambiguity and orbit inaccuracies
it is not possible to derive from the geometry. However, the relation between
changes in and is easily found to be

where the initial value of is obtained for the reference surface. Combining the
physical phase observations in eq. (2.4.5) with the geometric configuration expressed
in eq. (2.4.7) the relation between an interferometric phase change and the change
in the look angle is found to be
 2.4 SAR interferometry 37

The interferometric phase change can be defined as the difference between the mea-
sured (unwrapped) phase and the expected phase for the reference body derived
from the orbit geometry, hence

The height of the satellite above the reference body is known, and can be expressed
as

and the derivative for a resolution cell P with range gives the relationship
between a change in look angle due to a height difference

Note that this definition is independent of the choice of the reference body, although
fig. 2.10B is sketched for a flat earth. Furthermore it has to be noted, cf. fig. 2.10A,
that is the measured height for range resolution cell P, since P and P' are in
the same resolution cell. Using eqs. (2.4.8) and (2.4.11) we derive the relationship
between the height above the reference body and the phase difference

with

see fig. 2.10B. The initial value is found for an arbitrary reference surface (e.g., a
sphere or ellipsoid). A recursive scheme is used to find new values for at a specific
height above this reference surface. Inserting in eq. (2.4.12) yields the
height ambiguity—the height difference corresponding with a phase shift:

To conclude this evaluation, we combine the influence of topography, and surface

displacement, on the interferometric phase differences, relative to the reference
body. Using eqs. (2.4.5), (2.4.7) and (2.4.12), we find

Since the measured interferometric phase is the sum of the reference phase and
the deviations:
38 Chapter 2: Radar system theory and interferometric processing

where the reference phase is defined by

we find

Using ERS parameters, this implies that for an effective baseline of 100 meters,
a height difference of 1 meter yields an interferometric phase difference of ap-
proximately 4.5 degrees, which is well below the noise level of some 40 degrees, and
is therefore practically undetectable. However, in the differential case, a change
of 1 cm in the range direction, yields a phase difference of 127 degrees, which is
easily detectable.

2.4.2 Differential interferometry

Differential interferometry aims at the measurement of ground deformation using

repeat-pass interferometry. Since line-of-sight displacements enter directly into the
interferogram, independent of the baseline, it can be measured as a fraction of the
wavelength. Unfortunately, non-zero baselines will always cause some sensitivity
w.r.t. topography in the interferogram. The length of the perpendicular baseline
deviates from zero since (i) it is often not possible—or not desirable—to manoeuvre
the satellite in a zero-baseline orbit, and (ii), the baseline varies with the look angle,
yielding only one zero-baseline range-bin in the interferogram. Of course it depends
on the range of topography in the scene whether the baseline is significant.
There are several ways to construct a differential interferogram. The two-pass
method uses an external elevation model that is converted into radar coordinates,
scaled using the baseline, and subtracted from the interferogram (Massonnet et al.,
1993). Since for many areas in the world (crude) elevation models are available this
can be a feasible approach. Of course, errors in the DEM will propagate into the
deformation results, depending on the baseline characteristics.
A second method is the so-called three-pass method (Zebker et al., 1994b), see sec-
tion 3.5.1. Here, an extra SAR acquisition is used and combined with a suitable
partner acquisition to create the so-called topographic pair. This pair is assumed
to have no deformation, a suitable baseline providing sensitivity to topography, and
sufficient coherence. This pair is unwrapped, scaled to the baseline characteristics of
the deformation pair, and subtracted from it, yielding the differential interferogram,
the so-called differential pair.
In order to perform the three-pass method, using a topographic pair derived from
SAR image A and B and a deformation pair from image A and C, the effect of
the baseline difference between the two pairs should be taken into account. The
baselines can be scaled after the reference phase is subtracted from both interfero-
grams (Zebker et al., 1994b).
 2.4 SAR interferometry 39

The three-pass method can only be applied when both the topographic pair and the
deformation pair have a common image. This image is then used as a reference to
align the other two. Due to the available baselines for a specific scene, it might be
that the topographic pair and the deformation pair cannot be chosen with a com-
mon image. For example, the baselines for all candidate partners of the deformation
pair are too large, reducing coherence in the topographic pair and hampering phase
unwrapping. In that case, the four-pass method can be used, where the topographic
pair and the deformation pair are independent, i.e. they share no common SAR
acquisition, see section 3.5.2. As long as sufficient alignment of the two interfero-
grams can be performed the methodology is comparable to the three-pass method.
Practical differences with the three-pass method are manifested in the influence of
atmospheric signal in the three or four scenes and in the alignment.
Regarding the atmospheric signal, the three-pass method has one identical atmo-
spheric contribution for an arbitrary pixel in the common image. Since the topo-
graphic pair is scaled up or scaled down with scaling factor to match
the baseline characteristics of the deformation pair, also the atmospheric signal in
the topographic pair is scaled up or down. The influence of the scaling needs to
be considered carefully in the interpretation of the results. In section 3.5.1, the
functional model for differential interferometry is discussed in more detail.

2.4.3 Conditions for interferometry

SAR interferometry only works under coherent conditions, where the received re-
flections are correlated between the two SAR images. Evidently, this is the most
important condition for interferometry. Loss of coherence, known as decorrelation,
can be due to a number of driving mechanisms. The effect of some sources of decor-
relation, e.g., as introduced by the alignment and interpolation of the images, can
be reduced by using well-designed filtering procedures. Other sources of decorrela-
tion are more significant and non-reversible. The two most important conditions are
related to the phase gradient and the temporal variation in the physical distribution
of the elementary scatterers.
The phase gradient condition can be conveniently described in the spectral domain.
The temporal bandwidth of the SAR images in range corresponds with a spatial
bandwidth due to the projection on the earth’s surface. A phase gradient in range
of cycles/pixel corresponds with a spectral shift between the spectra of both ac-
quisitions of Hz, where is the sampling frequency. The spectral shift results
in a decreased overlap between the corresponding parts of the spectrum (the signal)
and an increasing non-overlapping part of the spectrum (the noise). Due to the
limited bandwidth, a phase gradient larger than cycles/pixel (approximately
0.822 for ERS) results in a zero overlap between the spectra, hence a complete loss
of correlation. The occurrence of this situation is dependent on: the length of the
perpendicular baseline, the steepness of the topographic slopes, and/or the gradient
of the surface deformation.
If topography w.r.t. an ellipsoidal reference surface is neglected, loss of correlation
due to the length of the perpendicular baseline occurs at approximately 1 km for
40 Chapter 2: Radar system theory and interferometric processing

ERS, commonly referred to as the critical baseline, A topographic slope

towards the sensor decreases the critical baseline locally. The relation between the
fringe frequency [cycles/m], the perpendicular baseline the incidence angle
and the local terrain slope (positive for slopes tilted towards the sensor) can
be expressed by (Gatelli et al., 1994; Bamler and Hartl, 1998):

Using the critical phase gradient of 0.882 cycles/pixel

we find a critical baseline of ~1100 m. Equivalently, a phase gradient due to surface
deformation in range direction needs to be less than 0.822 cycles/pixel in the radar
line of sight, corresponding to per 7.9 m or 2.9 mm/m. This gradient
of approximately forms an important upper limit for detecting slopes or defor-
mation gradients in the interferogram. It excludes, e.g., areas close to an earthquake
faulting area, where deformation gradients above may occur.
Figure 2.11, modified from Massonnet and Feigl (1998), gives an overview of the lim-
itations of the ERS SAR system, regarding deformation studies, although the change
in range could also be due to topographic slopes. The detectable combinations of a
change in range and the width of the phenomenon are limited within the five bold
lines, indicated in the figure. The upper deformation gradient of is shown,
as well as a lower deformation gradient of or 1 cm per 100 km. The lower
gradient is mainly due to orbit inaccuracies or long wavelength atmospheric
gradients.
The limits to the left-hand and right-hand side of the white area are given by the
physical limitations of the system, specifically the resolution and the swath width,
respectively. Finally the thermal noise of the ERS-AMI system imposes an upper
limit on the phase accuracy of about 12°, or l/30th of a fringe, which is equivalent
to 1 mm in range between any two pixels.
After identifying these conditions for the application of interferometry, we can eval-
uate whether specific geophysical phenomena comply with these conditions. For
example, events such as the rupture near an earthquake fault, or strong volcanic
deformations can easily exceed the limit. On the lower end of the plot, events
such as earth tides, loading, or postglacial rebound are hard to detect since they
have too long spatial wavelengths with respect to the imaging system and the
gradient limit. Of course, swaths including several consecutive images might be used
to circumvent some of these problems.
Some of the case studies reported in chapter 5 are indicated in fig. 2.11. For example,
the vertical movement of a corner reflector is at the edge of the range of detectable
deformations, with a width of only a few pixels and a change in range of less than
one-half phase cycle. For subsidence phenomena, as in the Cerro Prieto Geothermal
Field (CPGF) or the Groningen gas field, it is evident that these phenomena have a
strong time contraint. With a maximum subsidence rate of ~7 mm/yr and a spatial
extent of ~20 km, the deformation could become marginally detectable after 1 year,
although a time span of 5 years is more suitable for this phenomenon. On the other
 2.4 SAR interferometry 41

hand, subsidence rates at CPGF of ~1 cm/month over an area of only a few km

across are much easier to measure.
Temporal decorrelation, not indicated in fig. 2.11, is the second major limitation for
the application of repeat-pass interferometry, and a key problem in large parts of the
world when using large time intervals. For temporal decorrelation to occur, the in-
coherent sum of the variable scatterers within a pixel should be a significant fraction
of the coherent sum of stable scatterers. Weathering, vegetation, or anthropogenic
activities are common causes of temporal decorrelation.
A decorrelation effect can be directly related to the variance of the interferometric
phase and propagates to the parameters that are inferred from this phase. This is
discussed in depth in chapters 3 and 4.
Apart from error sources causing direct decorrelation, which is often easily observable
as noise in the interferogram, there are often limitations in distinguishing between
the different geophysical or induced sources of coherent phase variation. As stated
previously, the observed interferometric phase is a coherent superposition of effects
42 Chapter 2: Radar system theory and interferometric processing

from topographic height, surface deformation between the acquisitions, the state
of the atmosphere during both acquisitions, and possible errors in the orbit of the
sensor. The interpretation of a SAR interferogram, aimed at only one of these ap-
plications, is therefore dependent on prior knowledge about the other contributions.
For instance, in two-pass differential interferometry, an error in the a priori elevation
model yields artifacts in the deformation maps.
Other limiting conditions can occur in the interferometric processing sequence. Espe-
cially the process of phase unwrapping, in which the phase ambiguities are resolved,
can lead to local or global errors in the inferred height or deformation maps.
Other limitations in the interpretation of interferometry depend on the application
of the products. For example, for DEM generation, micro-topography may have to
be considered. Within each resolution cell the actual height may vary considerably,
for example by buildings. It depends on the dominant scatterers within the res-
olution cell, i.e., a roof top, what the resulting phase will be and therefore, which
height will be measured. In layover areas, scattering from various separated locations
may contribute to the phase observation, cf. fig. 2.9, leading to an ambiguity that
cannot be solved for. For areas that lack significant return signal (low backscatter
areas, i.e., due to specular2 reflection) the phasors are too short to provide a useful
phase observation. Finally, strong scatterers may create significant sidelobes in the
interferogram. These sidelobes can contaminate the phase observations in the neigh-
boring resolution cells significantly, which can result in an erroneous interpretation
of these phase values.

2.5 Interferometric processing overview

This section gives an overview of the sequence of most important decisions and
processing steps to create interferograms and geocoded products from SAR data.
The interferometric processing can start with focused complex SAR data, sometimes
referred to as Single-Look Complex (SLC) data, which may be available as a product
from the agency exploiting the satellite. Nevertheless, raw (unfocused) SAR data are
often preferred over SLC data, since they are usually cheaper, can be delivered faster,
and exclude the possibility of different focusing strategies at the various processing
facilities. In the processing steps described in the sequel, it doesn’t matter whether
image focusing, as described in section 2.3, has been performed by a third party
or if it is included in the interferometric processing chain. The only exception is
the azimuth filtering, which is not necessary using raw data since the mean of the
Doppler centroid frequencies of both images can be used in image focusing. For
SLC data, it is not known beforehand for which interferometric combinations the
data will be used. Therefore, the data are focused with respect to the zero-Doppler
frequency. To suppress noise introduced by the non-overlapping parts of the image
spectra, azimuth filtering may be required to remove these parts.

2
Specular reflection is the redirection of an electromagnetic wave at an equal but opposite angle,
as described by Snell’s Law in optics. Hence, the scattered signal is directed away from the radar
(Raney, 1998).
 2.5 Interferometric processing overview 43

2.5.1 Image selection

Image selection is perhaps one of the most vital decisions in the application of radar
interferometry, assuming that several SAR acquisitions are available. The criteria
depend on the specific application of the study. Main decisions regard the type
of sensor, the availability of the data, the temporal and spatial distribution of the
baselines, and the characteristics of the terrain and atmosphere during the image
acquisitions.
The sensor and platform characteristics include important parameters such as wave-
length, bandwidth, SNR, orbit inclination, and repeat period. Spaceborne repeat-
pass radar interferometry is mainly feasible from L-band to X-band
Below L-band, ionospheric signal will deteriorate the observations
and above X-band the instrument is too sensitive for the weather situation. The
wavelength will also influence the fringe density due to topography or deformation.
The sensitivity for topography or deformation is a combination between wavelength
and SNR. The range-bandwidth of the sensor determines the length of the critical
baseline and the range resolution. Platform characteristics such as orbit inclination
and the repeat interval determine the coverage of the earth, the occurrence of polar
gaps, and the revisit times between SAR acquisitions. Power considerations in com-
bination with other instruments on the platform can limit the operating time of the
sensor per orbit. Finally, the availability of precise tracking devices and the orbit
maintenance procedures influence the accuracy of the interferometric baseline.
The next concern after deciding on the type of sensor and platform is the data
availability. For satellites such as ERS, JERS, and Radarsat, convenient tools are
available on-line to browse through the archive of acquired images. Due to the
overlap between the SAR swaths of neighboring orbits, as well as the availability
of ascending as well as descending orbits, a certain area on earth may be imaged
from different viewing geometries. This enables the reduction of image distortion
effects and resolving different components of deformation vectors. Note that it is
not possible to create interferometric pairs from SAR images acquired from different
orbital tracks. Due to the maximum baseline restrictions, the viewing directions of
the two acquisitions should differ less than 4 minutes of arc for ERS. For scientific
satellites such as the ones listed above, it is possible to request future acquisitions
of a certain area, using a specific imaging mode.
Focusing on archived data, the distribution of temporal and spatial baselines be-
tween SAR acquisitions can be used to decide on the feasibility of interferometry
for a specific application. In order to obtain a quick overview of the possibilities, a
graphical representation of the spatial and temporal baselines can be convenient, see
fig. 2.12. The spatial (perpendicular) baselines give an indication of the sensitivity to
topographic height, the amount of decorrelation due to the phase gradients, and the
effectiveness of the phase unwrapping. Depending on the roughness of the terrain,
convenient perpendicular baselines for topographic mapping range between 100 and
500 m. The temporal baselines for topographic mapping should ideally be as short
as possible, to minimize the effect of temporal decorrelation. For interferograms
spanning surface deformation, a minimal perpendicular baseline is preferred to re-
44 Chapter 2: Radar system theory and interferometric processing

duce phase signal due to topography and noise due to the phase gradients. The ideal
temporal baseline depends on whether the deformation is expected to be continuous
in time or instantaneous, as e.g. earthquake deformation. For continuous deforma-
tion processes, the expected deformation rate should be used to choose temporal
baselines which ensure a sufficient signal-to-noise ratio between the deformation and
all other effects in the interferogram (noise, atmospheric signal, residual topographic
signal). For instantaneous processes, a short temporal baseline spanning the event
is usually ideal.
Some characteristics of the terrain are important to consider during the image se-
lection process. Relevant characteristics are roughness, elevation range, type and
amount of vegetation, and the influence of anthropogenic activities. The first two
terrain characteristics are evaluated in combination with possible baselines and sen-
sors, and may lead to the combination of ascending and descending interferograms.
The latter two characteristics give an qualitative indication of the amount of decor-
relation that can be expected. For example, on coarse scales (> 100 km) the monthly
Normalized Differential Vegetation Index (NDVI) as derived from NOAA AVHRR
observations can be used to approximate the amount of decorrelation due to vegeta-
tion in tandem imagery, see van der Kooij (1999b). To avoid seasonal effects, such
as snow or deciduous trees, it can be advantageous to select data from the same time
 2.5 Interferometric processing overview 45

of year. If surface deformation is the goal of the study, it is important to consider

the type, scale, and amount of deformation to be expected.
The amount of atmospheric signal to be expected during the acquisitions can be a
final consideration in selecting the data. Information about the state of the atmo-
sphere can be derived from satellite imagery or ground-based synoptic meteorological
data. Especially indications of convectivity in the troposphere, often related with
vertical refractivity structures, can be an indication for significant atmospheric signal
in the acquisitions. Likewise, choosing night-time instead of day-time acquisitions
will reduce the likelihood of atmospheric signal.
The considerations listed above may even lead to an optimal selection of potential
acquisitions and interferometric combinations. However, they may also lead to the
conclusion that radar interferometry is not the most suitable geodetic tool to fulfill
the goals of the study. Either way, proper consideration and evaluation of the options
is of major importance.

2.5.2 Preprocessing

If data processing starts from raw data they need to be checked for inconsistencies,
such as missing lines or sampling window start time (SWST) changes, before focusing
the images. If the number of missing lines is limited, the good lines heading or trailing
the inconsistencies can be copied to fill the gaps. This will only slightly deteriorate
image focusing. SWST changes within the image can be accounted for by aligning
both sides.
For interferometric processing, the spectra of the files should have maximum overlap
in azimuth direction. For pre-processed SLC data, where the desired interferometric
combinations are not known beforehand, an azimuth filtering procedure might be
necessary, see section 2.5.5. When raw SAR data are used, and the interferometric
combinations are known in advance, maximum spectral overlap in azimuth direction
can be obtained by estimating the mean Doppler centroid frequency, in each
image. Since determines the center frequency for the azimuth window used by
the SAR processor, there should be maximum overlap between the spectra of both
images. This can be obtained by using the average value of the two estimates.

2.5.3 Coregistration

The sub-pixel registration of both focused SAR images is a strict requirement for
interferometric processing. Depending on the start-stop times of a particular section
in the orbits (corresponding with the operation of the SAR), there can be an along-
track shift between the two SAR images, up to several thousands of lines. The
across-track shift is equal to the length of the parallel baseline divided by the posting
in range which can amount up to tens of pixels.
Since cross-correlation techniques for optimal alignment tend to be slow for very
large search windows, the procedure is usually separated in two steps: coarse and
fine coregistration. In the coarse coregistration, the offsets are approximated either
by defining common points in the image by visual inspection or by using the satellite
46 Chapter 2: Radar system theory and interferometric processing

orbits and timing as a reference. The subsequent fine coregistration usually applies
automatic correlation techniques to obtain sub-pixel alignment accuracy.
Coarse registration
The raw SAR data from ERS are tagged by a counter using a 3.4 ms interval. With
a relative velocity of 7.1 km/s on the ground, this corresponds with approximately
25 m in azimuth direction. Using precise orbit information the position of a point
at a certain range/azimuth position can be determined within some 25–50 m.
Such an approach usually applies two procedures. First, for an arbitrary point on
the master’s orbit and an arbitrary range the position of a pixel on the ellipsoid is
determined. Next, given this position an iterative algorithm searches along the slave
orbit until the correct Doppler position is found from which this pixel is observed.
Both orbit positions (in the master and the slave orbit) correspond to an exact
acquisition time with respect to their images start times (the acquisition times of the
first azimuth line) and using the pulse repetition frequency the azimuth line numbers
can be computed for both images. The along-track shift is the difference between
these line numbers. In range direction the distances to the common pixel, divided
by the range posting, yields the across-track shift. The procedure is comparable to
the geocoding of the interferogram, described in more detail in section 2.5.13.
Fine registration
Once the relative shifts between the two SAR images are determined within tens
of pixels in azimuth and a few pixels in range, the fine (sub-pixel) registration can
be performed. Just and Bamler (1994) have shown (for distributed scatterers) that
coregistration to an accuracy of l/8th of a pixel yields an almost negligible (4%)
decrease in coherence, see also section 4.4.6, as long as the data are acquired with
relatively small squint angles. Coherent registration techniques apply the full com-
plex (amplitude and phase) data to perform a complex cross correlation. These
techniques can be very accurate, but are generally not very robust. For situations
with large effective baselines, where many fringes appear in the estimation window
due to the side-looking geometry and topography, or with poor phase fidelity these
methods may fail. Incoherent techniques apply either only the (squared) amplitude
of the signal or even only the phase.
The cross-correlation of the powers (squared amplitudes) of the two images is a
commonly applied method for coregistration. The correlation peak indicates the
offset vector between the images in range and azimuth direction. For sufficiently
correlated images the offsets can be determined with an accuracy of l/20th of a
single-look pixel, corresponding with 20 cm in azimuth and 1 m in ground-range
for ERS. Note that since the cross-correlation product has a doubled bandwidth
compared with the two SLC images, aliasing will be introduced. To avoid this, both
datasets need to be oversampled by at least a factor two before calculating the cross-
correlation. Prati and Rocca (1990) oversample a set of windows of 100 x 100 pixels
in both complex images by a factor of 8 using an FFT interpolation. The relative
shift between the two images is then determined for every window.
Once the relative registration shifts are determined for a set of windows, ideally
 2.5 Interferometric processing overview 47

evenly distributed over the whole scene, the shifts can be regarded as a set of dis-
placement vectors to restrict a two-dimensional polynomial to determine the dis-
placement vectors for every single pixel in the slave image. The polynomial trans-
formation might be up to a fifth degree, although usually offsets and stretch param-
eters in both directions are sufficient, comparable to a bi-linear model. See Brown
(1992) for a general review paper on coregistration and Michel et al. (1999) for an
applied overview. Coregistration for SAR interferometry is discussed in Gabriel and
Goldstein (1988); Gabriel et al. (1989); Prati et al. (1989); Prati and Rocca (1990);
Weydahl (1991); Lin et al. (1992); Hartl and Xia (1993); Zebker et al. (1994c);
Geudtner et al. (1994); Geudtner (1995); Carrasco et al. (1995); Coulson (1995);
Homer and Longstaff (1995); Schwäbisch and Geudtner (1995), and Samson (1996).

2.5.4 Resampling and interpolation

The interferometric combination of the two complex images requires evaluation of

the complex values in one of the two at the pixel positions of the other. Although
the implementation may be different, this resampling can be regarded as consisting
of two subsequent steps: (i) reconstruction of the continuous signal from its sampled
version by convolution with an interpolation kernel, and (ii) sampling of the recon-
structed signal at the new sampling locations. The choice of the interpolator kernel
(especially its length) requires a tradeoff between interpolation accuracy and com-
putational effort. Using two subsequent one-dimensional interpolation algorithms,
the amount of floating point operations is in the order of for an interpolation
kernel with points. Kernel sizes vary between 1 (nearest neighbor), 2 (linear), and
4, 8, or 16 for extended kernels. The corresponding interpolation times are in the
order of and respectively. In section 4.4.5, it is shown
that 4-point or 6-point kernels are usually sufficiently accurate. Moreover, for larger
baselines spectral shift filtering in range increases the oversampling ratio so that even
2-point kernels may be of sufficient quality. The same holds in azimuth direction, if
the Doppler centroids are different, as may occur using ERS-1 and ERS-2 combina-
tions. For high-quality interferogram processing, the complex data are oversampled
by a factor of 2 before the complex multiplication of both datasets, to avoid aliasing
effects. Then, linear (2-point) interpolation is sufficiently accurate, see fig. 4.11.
The interpolator kernel is centered at zero-frequency. Therefore, the suppression of
the spectrum of the signal is most severe for the higher frequency. As the signal
spectrum of SAR data is only centered in range direction and not in azimuth, a
Doppler centroid frequency that deviates significantly from zero-Doppler will be
distorted non-symmetrically. To avoid this, it is recommended to shift the signal
temporary to the zero-Doppler frequency, perform the interpolation, and shift the
interpolated signal back to the original Doppler centroid frequency.

2.5.5 A priori filtering

Phase noise in the interferogram can lead to problems in the phase unwrapping
(high number of residues) or hamper data interpretation. Although the effects in
the interferogram might be similar, the driving mechanisms behind the noise may
48 Chapter 2: Radar system theory and interferometric processing

be different and require a different approach. To account for this difference, we

distinguish between a priori filtering and a posteriori filtering. A priori filtering is
applied to the original SAR data, before interferogram generation, whereas a posteri-
ori filtering is applied after the interferogram has been formed—if applicable even to
stacks of interferograms. A posteriori filtering is discussed in section 2.5.8. Complex
multilooking is a filter operation commonly applied during interferogram formation.
The main characteristic of the error sources attacked by a priori filtering is the sep-
arability of signal and noise in the spectral domain. Consider two image spectra
that contain a distinct signal part and noise part, as in fig 2.13. The complex multi-
plication of the two images, as applied during interferogram formation, corresponds
with a convolution of the two spectra. In the convolution, the noise parts of the two
original images will get smeared over all frequencies in the spectrum of the resulting
interferogram. Since the noise components cannot be effectively separated at this
stage, a posteriori filtering of the interferogram for this type of noise is not very
effective.
Spectral shift compensation
Spectral shift compensation aims at increasing the SNR of the product by eliminating
the noise components in the two datasets (Gatelli et al., 1994; Prati and Rocca,
1994b; Just and Bamler, 1994). Increasing the SNR improves coherence, hence
better phase statistics and less problems during phase unwrapping. Slightly different
approaches need to be applied in range and azimuth direction.
The origin of the spectral shift problem is the difference between the object spectrum
and the data spectrum, as sketched in fig. 2.14A. The object spectrum is a function
of the object properties, in this case the radar reflectivity of the terrain, and is
independent of the imaging system. In contrast, the data spectrum reveals the
characteristics of the imaging system—its bandwidth contains the measurable width
of the spectrum, and its Nyquist frequency is determined by the system’s sampling
frequency. Ideally the object spectra should be mapped identically to the data
spectra in order to apply multiplicative interferometry.
A spectral shift is caused by the difference in the local incidence angles towards the
two sensors, see fig. 2.14B. If both sensors would be at exactly the same position and
look in the same direction, no phase-differences would occur in the image, hence,
 2.5 Interferometric processing overview 49

no fringes. Since the time-domain (the image) will be equal for both sensors, the
spectral domain is equal as well. Consider now a different geometry, with sensor 2
at a slightly different position from sensor 1. Now phase differences—fringes—will
occur in the image. Albeit the location of the data spectrum is indifferent (the data
do not sense a difference in the local incidence angle), its contents have shifted to
other frequencies, cf. fig. 2.14A. In other words, the ground-to-slantrange mapping
causes a wavenumber shift—the wavenumbers of the object spectrum have shifted
to other frequencies in the data spectrum.
The problem of the wavenumber shift is that both images have an overlapping spec-
tral part, which contains information, and two non-overlapping parts, which can be
considered as noise for interferometric purposes. Therefore, a bandpass filter needs
to be applied to eliminate the two parts containing noise. To tune the bandpass filter
it is necessary to determine the wavenumber shift as accurate as possible. The fringe
frequency reflects the amount of spectral shift and hence the amount of decorrelation
due to the non-overlapping parts of both spectra. It can be estimated locally from
the data—as long as there is sufficient correlation between the two datasets—or from
the orbit state vector information. In the latter case, the change in range between
near range and far range is used to compute the average phase gradient (fringe fre-
quency), corresponding to one spectral shift value for the entire image. However, for
topography with a variety of slopes the filtering can be too coarse—resulting in
an unnecessary loss of resolution—or not significant enough—resulting in an unnec-
essary loss of coherence. In the former case, the filtering can be performed per range
line, or even in small patches to adapt its performance to local slopes. The patches
are complex multiplied to form an interferogram strip which is Fourier transformed.
The fringe frequency follows from the peak in the amplitude spectrum.
The fringe frequency is related to the perpendicular baseline the incidence angle
and the local slope of the terrain in the zero-Doppler direction, cf. eq. (2.4.19)
50 Chapter 2: Radar system theory and interferometric processing

(Gatelli et al., 1994; Bamler and Hartl, 1998):

where is the slant-range to the reference satellite. After the spectral weighting
(Hamming in the case of ERS) is removed, the non-overlapping parts of the spec-
tra of both datasets can be set to zero-values, and the Hamming window can be
applied again to the newly formed spectra, while considering the new bandwidths
The smaller bandwidth translates in a reduced resolution in range
direction, see eq. (2.3.9). Therefore, range filtering can yield a significant increase in
correlation between the two images at the expense of a reduced resolution (Gatelli
et al., 1994).
Spectral shift in azimuth direction
The procedure in the azimuth direction depends on the input products for the in-
terferometric processing. Working with raw radar data, the Doppler centroid fre-
quencies and of the two datasets can be estimated, see section 2.3.4. By
processing both datasets at the average (center) value, the relative spectral shift
can be accounted for. Note, however, that this approach does not account for the
weighting function, cf. fig. 2.4, on the antenna pattern (Schwäbisch, 1995a).
SLC data have been processed at a fixed Doppler Centroid frequency, and as a result
tuning the processing for a specific interferogram combination is not possible
anymore. In this case, both spectra will appear shifted in the SLC data spectrum.
Similarly as in range direction, the non-overlapping parts of the spectrum will result
in noise in the interferogram.
The difference in between both images can be caused by (i) the different squint
angles of the sensors, or (ii) the convergence of the orbits. Both possibilities have
the same effect: a point on the terrain is viewed from two different positions. The
different imaging geometry causes a different mapping of ground-range to slant-
range, therefore an increased/decreased sensitivity for e.g. higher frequencies, which
is translated to a shift in the spectral domain of the surface within the data spectrums
bandwidth.
Bandpass azimuth filtering can be performed in order to remove the non-overlapping
parts of the azimuth image spectra. At the same time the oversampling by a factor
of two can be performed, so the output data are twice as large as the input data.
These operations are performed in the frequency domain. Both images need to be
filtered. The weighting function used by the SAR processor needs to be selected
beforehand (often Hamming weighting)
Azimuth filtering may be necessary whenever the Doppler centroid frequencies of
the two SAR images differ too much. This is a result of a difference in the squint
angles of the radar antennas. An effective squint angle can be introduced by the
earth’s rotation. For example, the orientation of the antenna beam of Seasat was
orthogonal to its orbital plane, causing ±3° squint variation for each orbit due to
the earth’s rotation. The phased array antennas of the ERS systems, however,
apply a technique called “yaw-steering,” which largely corrects for earth rotation
 2.5 Interferometric processing overview 51

effects. Effectively, the antenna beam is maintained orthogonal to the spacecraft’s

groundtrack, which translates in a symmetry with respect to zero-Doppler within
a tolerance of 5% (Henderson and Lewis, 1998). For this reason, effective azimuth
squint angle variation for ERS is limited.
Doppler centroid frequency estimation for SLC images
When SLC images are used as input, the Doppler centroid frequency variation
(mainly over range) can be estimated to obtain the input values for azimuth filter-
ing. The frequency can be estimated over a strip of SLC imagery, small in azimuth
but wide in range using a discrete Fourier transform. Areas with homogeneous in-
tensity or water areas are particularly suitable. Although these estimates are not
very accurate they are often good enough for locating the zero-padding location (for
oversampling) and coregistration.

2.5.6 Interferogram formation

A complex interferogram is constructed by a pointwise complex multiplication of

corresponding pixels in both datasets, see eqs. (2.4.1) and (2.4.2):

Although especially the phase of pixel P is of importance for in-

terferometry, the complex interferogram is usually retained for future computations.
Apart from the fact that the amplitude of the phasor contains important information
on the SNR of the phase observations, see fig. 2.15, the complex representation does
not have the artificial phase jumps. This is convenient, e.g., when orbital phase
gradients need to be subtracted from the interferogram. Furthermore, a posteriori
multilooking to improve phase variance cannot be applied with optimal accuracy
when the amplitude information is discarded.
52 Chapter 2: Radar system theory and interferometric processing

Oversampling
The complex multiplication of two aligned and interpolated SLC datasets corre-
sponds with a convolution in the frequency domain. Consequently, the resulting
interferogram will have a doubled bandwidth. Although the raw radar data have
been truncated to a bandwidth smaller than the sampling frequency, both in range
and in azimuth, the convolution of both datasets will yield spectral contributions
above the Nyquist frequency, which results in aliasing effects, disturbing especially
the short wavelength signal in the interferogram. To avoid aliasing effects in this part
of the spectrum, the data are oversampled by a factor 2 (or possibly less whenever
the bandwidth is already reduced due to azimuth/range filtering) before evaluating
the complex multiplication. The resulting interferogram is usually low-pass filtered
to obtain the same resolution as the SLC images. In this approach aliasing effects
are eliminated.
Taking looks
To decrease the noise in the interferogram, Goldstein et al. (1988) propose a multilook
approach, in which the complex interferogram data in a specified window are simply
averaged. Such a boxcar convolution approach corresponds with a two-dimensional
sinc multiplication in the spectral domain. In section 4.2.2 it is shown analytically
that the PDF of the interferometric phase improves in the multilook case (Lee et al.,
1994). Multilooking can be performed simultaneously with the complex multiplica-
tion, and is often applied to a range-azimuth ratio that yields approximately square
pixels, such as 1 : 5, 2 : 10, etc. The procedure corresponds to boxcar filtering fol-
lowed by subsampling. Note that the term multilooking strictly has another meaning
in SAR processing, as discussed in section 2.3.7.

2.5.7 Reference phase

Using the satellite state vectors, the orbits of both satellites are being modeled by
a polynomial of the first or second degree during the observation period of the two
images. To determine the expected phase behavior for a reference body, the distances
between the orbits and the reference body need to be evaluated and differenced for
every resolution cell. Standard reference bodies are global ellipsoids such as WGS84
or locally best-fitting ellipsoids such as Bessel. The procedure can be summarized
in four steps:

1. at a few positions (times) along orbit 1, the ranges to a few equally distributed
points in the interferogram area are determined,

2. for these points, the position at the time of imaging along orbit 2 is retrieved
together with the corresponding ranges,

3. the range differences are determined for every reference point, converted to
phase differences, and

4. interpolated (or evaluated) for every resolution cell of the interferogram.

 2.5 Interferometric processing overview 53

In the first step, the Doppler frequency used for the SAR focusing, the range time,
and the definition of the reference body are used to find the location of point P, by
solving a set of three equations (Schwäbisch, 1995a).

The point is effectively viewed from an angle defined by the processing Doppler
frequency (zero-Doppler for ESA SLC products), cf. eq. (2.3.13),

where is the velocity vector, the squint angle, and and are the range
vector and its derivative in time respectively.

Point P is located on the reference body.

The observed range corresponds with the difference between the state vector
and the positioning vector of the reference point.

These three equations can be solved iteratively. The satellite state vectors are given
in a geocentric earth-fixed coordinate system, such as the Conventional Terrestrial
Reference System (CTRS), and the reference surface needs to be known in the same
system.
In the second step the range from point P to the second orbit is evaluated, under the
condition that this location satisfies the Doppler centroid mapping for the second
image. The third and fourth step are straightforward.
Note that in some cases the satellite orbits are not known or not known with sufficient
accuracy. In those cases, the baseline can be approximated from the data, especially
the range offset vectors (a measure for the parallel baseline) and the fringe frequency
(which is a function of the perpendicular baseline) (Zebker et al., 1994b).

2.5.8 A posteriori filtering

Apart from complex multilooking during interferogram formation, or spectral shift
filtering prior to interferogram formation, other filters have been proposed to en-
hance the quality of the interferogram. Isotropic filters such as square boxcar filters
tend to perform well in flat areas, but fail in areas where the phase gradient is
high with respect to the kernel size. For example, if there are several phase cycles
within such a filter kernel, the resulting (filtered) phase will have a value which
does not necessarily reflect a “mean” value. Non-isotropic filters take into account
that the fringe frequency often has a dominant direction, and adjust the filter kernel
adaptively to the fringe frequency and slope.
Schwäbisch (1995b) uses a direction dependent filter that uses the local fringe fre-
quency for the orientation. The frequency domain shape of the lowpass filter is a
two-dimensional Gaussian, where the extent of the filter is highly non-isotropic to
suppress frequencies perpendicular to the dominant fringe frequency direction.
54 Chapter 2: Radar system theory and interferometric processing

Goldstein and Werner (1998) propose an adaptive filter, based on smoothing the
spectrum of the interferogram in patches using the amplitude of the spec-
trum. The spectrum is weighted by multiplication with its own (smoothed) intensity
to the power of an exponent

where is the spectrum of the filtered interferogram. Hence, for the

multiplication factor is one, and no filtering applies. However, for larger the
filtering will be significant. The methodology ensures that the signal in the spectrum
(assumed to have strong amplitudes) will be enhanced, whereas the noise (assumed
to have a significantly lower amplitude) is relatively suppressed. This way, the filter
is non-isotropic and changing to the characteristics of the terrain. For example, for
water areas, where all frequencies are assumed to have comparable amplitudes, the
filter will not enhance or suppress anything. On the other hand, a topographic slope
will have a distinguishable amplitude in the spectrum, which will be enhanced by
the filter. The filter is applied on overlapping patches in the interferogram. After
inverse Fourier transformation, they are reassembled with a linear taper in and
direction.

2.5.9 Phase unwrapping

For interferometric applications, the principal observation is the two-dimensional

relative phase signal, which is the of the (unknown) absolute phase
signal. Although the forward problem—wrapping the absolute phase to the
interval—is straightforward, the inverse problem is one of the main difficulties in
radar interferometry, essentially due to its inherent non-uniqueness and non-linearity,
see Ghiglia and Pritt (1998) for general reference. Variable phase noise, as well as
geometric problems such as foreshortening and layover are the main causes why many
proposed techniques do not perform as desired. Without additional information, the
assessment whether the results of phase-unwrapping are trustworthy will always be
based on strong assumptions on the data behavior.
The proposed methodology of analyzing interferometric data as a geodetic technique,
which will be described in chapters 3 and 4, is an attempt to approach the problem
of phase unwrapping as an adjustment and filtering problem. This approach is based
on the availability of more than one interferogram of the area under consideration.
First promising results using such an approach have recently been presented by
Ferretti et al. (1997). Here, for the sake of completeness, we briefly discuss the more
conventional algorithms and methods.
Let us denote the observed (wrapped) phase as a function of the unknown true
phase with

where W is the wrapping operator, is the difference in range to

the two satellite positions, expresses additive phase noise, and
 2.5 Interferometric processing overview 55

is the integer ambiguity number. This is largely following the notation of Bamler
and Hartl (1998). Without any assumption, it is not possible to solve the integer
ambiguity number Usually, based on a priori knowledge of the terrain, it is
assumed that the phase gradient between adjacent pixels is limited to the
interval; the smoothness criterion. Hence, if the observed phase gradient is larger
than a cycle is subtracted, and if it is less than a cycle is added. Now,
the phase gradients of the wrapped phases are assumed to be equal to the gradients
of the true phases. As long as the true phase gradients are small—less than one-
half cycle—and unaffected by noise, this assumption is valid. The actual phase
unwrapping is then performed by integrating the phase gradients, e.g., simply by
a one-dimensional (flood-fill) summation of the phase gradients, starting from an
arbitrary seed location. However, if one of the estimated gradients is erroneous,
this error will propagate throughout all subsequent pixels and may affect the entire
image.
The smoothness criterion implies that the true phase gradient field exhibits the
characteristics of a conservative potential field, i.e., the curl of the field is equal to
zero (Kellogg, 1929; Bamler and Hartl, 1998):

The gradient of the true phase is unknown, and has to be estimated based on
the gradient of the wrapped phase. For a specific pixel with coordinates
the best estimate is usually denoted by

Since noise in the phase data will yield errors, in general does not necessarily
equal resulting in global phase errors after integration.
The most basic evaluation of the smoothness criterion in a digital image is on 2 × 2
neighboring pixels, see fig. 2.16:

where the value of can be zero (no residue), +1-cycle (positive residue) or – 1-
cycle (negative residue) (Goldstein et al., 1988). The occurrence of residues in the
wrapped phase image is an indication of noise or undersampling and inconsistent
with the smoothness condition. It will yield global phase errors in the unwrapped
interferogram. Residues can be discharged by “connecting” nearby positive and
negative residues with each other, so that the total interferogram is unloaded. Using
this approach, true phase gradients that exceed one-half cycle can be identified and
accounted for.
Many new phase unwrapping algorithms have been proposed since SAR interferom-
etry became an active field of research. Here we briefly discuss some of the more
popular ones.
56 Chapter 2: Radar system theory and interferometric processing

Residue-cut method In the residue-cut method, also known as the branch-cut or

(minimum spanning) tree algorithm, all residues are connected (unloaded) with
branch cuts, under the condition that the total length of all cuts is minimal
(Goldstein et al., 1988), see also Arfken (1985) and Ching et al. (1992). The
algorithm starts at one residue and searches within a window of 3 × 3 around
it for another residue. If a residue of different sign is found, the branch cut is
applied and the algorithm continues to the next residue. If the found residue
is of equal sign, the search continues from this new position, and repeats until
all plus and minus residues are unloaded, hereby spanning a “tree” (Chen and
Zebker, 2000a). If there are no new residues within the search region, but the
found ones are not unloaded yet, the search radius is doubled to prevent the
algorithm from breaking off. While integrating the phase gradients, the cuts
constitute an obstacle through which the integration cannot continue. As a
result, isolated regions can occur in the interferogram, surrounded by cuts.
An important characteristic of the residue-cut algorithm is the fact that the
estimated phase gradients can only have discrete offsets with In
practice, many trees will form in noisy areas such as water surfaces, layover
regions, and regions affected by temporal decorrelation.

Least-Squares Method (Ghiglia and Romero, 1994, 1996; Pritt, 1996; Hunt, 1979)
These methods appear in a weighted and unweighted fashion, and are based
on minimizing the sum of the quadratic errors between the unwrapped phase
gradients and the estimated phase gradients in a global way. FFT techniques
can be applied to solve the least-squares technique in an efficient way (Ghiglia
and Romero, 1994). Considering the problem in a Green’s-function formula-
tion can also be regarded as a least-squares method (Fornaro et al., 1996a,b).
Bamler et al. (1996a) proved that least-squares methods for phase unwrapping
can introduce errors which propagate through the image, effectively by under-
estimating slopes (gradients) in the interferogram, see (Bamler et al., 1998).
 2.5 Interferometric processing overview 57

Minimal cost flow methods (Costantini, 1996, 1998; Flynn, 1997; Chen and Ze-
bker, 2000a) Minimal cost flow (MCF) or network flow methods consider the
phase unwrapping problem as a global minimization problem with integer vari-
ables. Comparing the difference between the estimated and the unknown phase
gradient of the unwrapped phase should yield differences, with
MCF methods follow the residue-cut methods, adding integer cycles to cross
the cuts. Positive and negative residues are the surplus and demand nodes,
and the “flow” that equalizes the nodes can travel along paths determined by
the cost function. An optimal solution is reached if the predefined objective
function is minimized. A practical approach for the definition of costs for
topography or deformation mapping is given in Chen and Zebker (2000b).

Linear integer combinations of interferograms, as discussed in section 3.5.3, is a

methodology to reduce the fringe rate, which can facilitate phase unwrapping (Mas-
sonnet et al., 1996c)

2.5.10 Patch unwrapping

Phase unwrapping methods applying branch-cuts, as well as interferograms contain-

ing, e.g., islands, long waterways, or joint areas of low coherence such as moving
lava flows are likely to contain isolated patches with sufficient coherence for success-
ful phase unwrapping, surrounded by incoherent areas or branch-cuts. The phase
gradient integration necessarily starts from an arbitrary seed, and can continue only
within the limits of the region surrounded by branch-cuts. After unwrapping as
many of these regions as possible, the connection of the separate patches is here
referred to as patch unwrapping.
Placing seeds—the starting points for phase gradient integration—in a dense grid
minimizes the possibilities that coherent patches are missed. Proper bookkeeping is
required to label the unwrapped patches and prevent unwrapping starting from a
seed in a patch that has been unwrapped already. The final output is an interfero-
gram, with separately unwrapped patches and a map with patch numbers. The task
of patch unwrapping is theoretically as ill-posed as phase unwrapping, unless some
sort of prior information on the terrain or the deformation pattern is added. For
example, for a short perpendicular baseline the height ambiguity might be so large
that an integer jump between patches is very unlikely. Often manual intervention is
necessary to apply the patch unwrapping. In section 5.4 two possibilities for patch
unwrapping are presented and demonstrated.

2.5.11 Differential interferogram generation

Producing a differential interferogram requires the two interferograms (the deforma-

tion pair and the topographic pair) to be aligned to the same grid. In three-pass
differential interferometry this requirement is satisfied by using a common reference
image, onto which the other two are coregistered and resampled. The four-pass
technique requires a repeated coregistration and interpolation. First the two inter-
ferograms are formed separately, followed by the coregistration of one of the pairs
58 Chapter 2: Radar system theory and interferometric processing

onto the other. After all images are in the same grid, the topographic pair needs to
be unwrapped and scaled according to the baseline ratio between the two interfero-
grams (Zebker et al., 1994b). Subtracting this unwrapped, scaled topographic pair
from the deformation pair yields the differential pair, which should ideally reflect
only deformation signal. The procedure is described mathematically in chapters 3
and 4.

2.5.12 Phase to height conversion

Phase to height conversion is the procedure that relates the unwrapped interfero-
metric phase to topographic height, although the pixels remain in the range-azimuth
radar coordinate system. The theoretical concept is already outlined in section 2.4.1,
but a few practical aspects of the implementation are important to consider.
First, a theoretical situation such as the one sketched in figs. 2.10A and 2.10B will
never occur, since it is not possible to derive the phase difference (and hence the
difference in look angle between a point on the ellipsoid and a point at the same
range with a different height, because of the wrapped nature of the interferometric
phase. Although this problem cannot be solved for a single pixel, the evaluation of
the unwrapped phase gradients to height gradients is straightforward. Therefore the
height gradients, or equivalently the relative heights, can be determined up to an
arbitrary constant for the entire image.
Equation (2.4.12) needs to be evaluated for the phase to height conversion. Fixed
parameters in this equation are and for the imaginary point P'
on the reference surface. However, the perpendicular baseline as well as the look
angle will change as soon as the initial height has been derived, see fig. 2.10B.
Therefore, the procedure has to be repeated iteratively until satisfies some ac-
curacy criterion. Usually not more than three iterations are necessary to obtain
mm-scale increments. Finally, the selection of one control point will fix the relative
height differences to the coordinate system of the reference surface.
The evaluation of the perpendicular baseline per pixel is vital, not only because of
the dependency on height. In fig. 2.17A, the variation of the baseline per pixel is
shown for non-parallel orbits. Orbit 1 observes all pixels in one range line from
one position, whereas the corresponding positions from orbit 2 are changing. For
the second range line this behavior repeats for a slightly shifted position, leading
to a sawtooth pattern. As a function of time (fast time and slow time combined),
the behavior of the perpendicular baseline is as indicated in fig. 2.17B, assuming a
reference surface without topography.

2.5.13 Geocoding

Geocoding refers to a coordinate transformation of the radar coordinates (range/-

azimuth/height) to coordinates in a convenient geodetic reference system such as
WGS84 where we use and for the geographic latitude and longitude,
respectively, and H for the height above the reference body (ellipsoid). For com-
putational convenience, an intermediate earth-fixed Cartesian reference system is
 2.5 Interferometric processing overview 59

used, in which the satellite orbits as well as the ellipsoid are given. The procedure
is similar to the evaluation of the reference phase, see section 2.5.7, and is based on
the following conditions.

1. For every pixel in the image, the range bin number is used to find the magnitude
of the range vector, to the reference orbit.
2. The azimuth position along this orbit is retrieved from the line number, the
azimuth starting time, and the PRF. This defines the state vector for
the satellite position.
3. The Doppler centroid, defining a possible squint angle, and the direction of
the local velocity vector of the state vector, are used to define a plane. The
plane is spanned by the origin of the coordinate system, the position of the
satellite, and the squint angle with respect to the velocity direction. The range
vector must be situated in this plane.
4. Finally, the local earth radius defined by the ellipsoid, plus the topographic
height derived from the interferometric phase, define the length of the
positioning vector

These conditions uniquely define the position vector in the geocentric Carte-
sian reference system. Latitude and longitude can easily be derived by a three-
dimensional inverse Helmert transformation.
Applying this procedure for every pixel in the image yields a long list of
triplets, where H will be assigned a not-a-number-value if the height could not
be derived, e.g., in decorrelated areas such as water, or in shadow or layover re-
gions. A non-zero height will result in a horizontal displacement of approximately
where is the incidence angle. As a rule of thumb, for ERS condi-
tions this horizontal displacement is approximately 2.3 × H. Therefore the positions
60 Chapter 2: Radar system theory and interferometric processing

of the points are irregularly distributed, and interpolation is necessary to

resample the data to a rectangular grid. Delaunay triangulation (Delaunay, 1934)
followed by linear interpolation is an often applied method in this respect.
An example of geocoding a SAR derived DEM is shown in fig. 2.18. The interfero-
gram (fig. 2.18A and B) covers an area of 16 × 10 km over central Java, Indonesia,
showing two volcanoes, Mt. Papandayan in the west and Mt. Cikuray in the east.
Clearly visible in the amplitude image are the foreshortening, layover, and shadow
areas. The interferogram has a perpendicular baseline of ~108 m, corresponding
with a height ambiguity of ~87 m. After a posteriori filtering using the method
of Goldstein and Werner (1998), phase unwrapping, and phase-to-height transfor-
mation, the data were geocoded to WGS84. Local tiepoints were used to remove
the bias in the height, to remove orbital errors, and to fine-tune the geocoding.
Figure 2.18C shows a 3D perspective of the derived DEM.
 Chapter 3
Functional model for radar interferometry

A generic Gauss-Markoff model is introduced and applied to describe the func-

tional and stochastic relations for two-pass spaceborne radar interferometry.
This chapter covers the functional part of this model, which describes the phys-
ical and geometric relations between the observations and the parameters. The
model is refined for specific applications such as topography estimation, defor-
mation mapping, and atmospheric mapping. An extended model is presented
to treat three-pass and four-pass differential interferometry, as well as inter-
ferogram stacking.
key words: Functional model, Topography, Deformation, Atmosphere

The interferometric processing relates two registered signals and to the Her-
mitian product which has a phase equal to the phase difference of the two
signals. The phase difference values for each pixel in the interferogram form the
observation space, which is related to the parameter space by a functional relation.
The following sections define a standard model to express this relation.

3.1 Gauss-Markoff model definition

The observations form a real stochastic vector of observations character-

ized by the first moment and the second moment After linearization,
the relation between the observations and the unknown parameters can be written
as a Gauss-Markoff model (Gauss, 1809; Markoff, 1912), see Koch (1999):

where A is the design matrix, is the variance-covariance matrix,

the real positive-definite cofactor matrix, and the a priori variance
factor. The vector of parameters is assumed to be real and non-stochastic.

The first part of eq. (3.1.1) is commonly referred to as the functional model or model
of observation equations, whereas the second part is known as the stochastic model.
62 Chapter 3: Functional model for radar interferometry

The latter is discussed in chapter 4. The functional model can also be written as

where reflects the errors of with Design matrix A should

be of full rank to make the model uniquely solvable, which implies that
Moreover, should be positive definite, since in that case the weight matrix of the
observations, will exist and will be positive definite as well (Koch, 1999). The
best (minumum variance) linear unbiased estimator of the unknown parameters
and the covariance matrix is given by

The advantage of describing the problem of parameter estimation in a systematic

model such as the one outlined above is that different physical or geometrical aspects
of the problem are reduced to relatively simple mathematical equations. Apart from
providing a structural and perhaps elucidating approach, the formulation in a Gauss-
Markoff model opens up a wealth of standard techniques for adjustment, quality
description, and hypothesis testing. As a result, such techniques give a quantitative
measure of the accuracy and reliability of the estimated parameters.

3.1.1 Observations

The primary observations are the raw radar echoes as received by the SAR antenna.
Unfortunately, using this definition it is very difficult to obtain the fully linearized
design matrix since SAR data processing is a highly non-linear operation (Bamler
and Hartl, 1998). To simplify the model, we therefore define the focused inter-
ferogram as observations, and use system theoretical considerations to describe the
stochasticity of the observations, see chapter 4. Some modifications of this definition
are necessary to arrive at eq. (3.1.1), since (i) the interferogram consists of complex
values, (ii) the interferometric phase is in first instance only known modulo and
(iii) the interferometric phase might be altered due to the extraction of a reference
and/or a topographic phase.1 In fact, several definitions for the observation vector
might be suitable, as long as the model in eq. (3.1.1) can be adopted. Here we
define four types of observation vectors, based on (i) whether or not the influence of
the reference body is accounted for, and (ii) whether or not the phase observations
are wrapped. The four types are listed in table 3.1, where is the wrapping
operator, with The reference phase is in first
approximation regarded to be a deterministic parameter defined by

1
Another option is to define the gradient of the interferometric phase as observations, see
Sandwell and Price (1998). This can be advantageous for identifying very localized gradients,
e.g., due to surface deformations, and in the analysis of stacks of interferograms. Here we choose
not to use this type of observation, as it relies on the assumption that the gradients are restricted
to a interval.
 3.1 Gauss-Markoff model definition 63

where and are the baseline length and orientation (see fig. 3.2) for resolution
cell These values are considered to be known, either by using precise orbit infor-
mation or after an adjustment procedure involving tie-points, see also section 4.6.3.
The look angle follows from the definition of the reference body in the same datum
as the orbits. The option to include the reference phase in the vector of observa-
tions, instead of including it in the design matrix, follows from its independence of
the parameters to be estimated, resulting in a rank deficiency. Note that the four
types of phase observation might be multilooked.
An interferogram of rows and columns consists of observations.
A random (stochastic) wrapped interferometric phase observation for pixel is writ-
ten as with where and represent
the row and column of the interferogram, respectively, see fig. 3.1. The vector of
observations is written as For the wrapped phase and for the
unwrapped (absolute) phase

3.1.2 Parameters and a priori information

Repeat-pass radar interferometry can be used for the recovery of geophysical pa-
rameters. Although many parameters affect the interferometric phase, at least five
parameters are dominant. For every observation these are

1. topographic height
64 Chapter 3: Functional model for radar interferometry

2. deformation in slant direction

3. slant-atmospheric delay during acquisition 1:
4. slant-atmospheric delay during acquisition 2: and
5. the integer ambiguity number

This results in a vector of parameters which is assumed to be real and non-

stochastic. This formulation implies that every observation is related to at least five
unknown parameters, which results in a rank defect of Therefore, the problem
is underdetermined. To solve this problem three strategies can be applied:

(i) more observations can be added to the model,

(ii) deterministic (a priori) information can be introduced, or
(iii) the model can be reformulated.

The first strategy is useful when a series of SAR acquisitions is available over the
area of interest. More than one interferometric pair can be processed, and assum-
ing that some parameters do not change for different interferograms the rank defect
decreases. The second option can be used whenever one or more parameters are
known, e.g., from different sources. A reference elevation model, a high likelihood
that no deformation occurred, or confidence in the phase unwrapping might be used
for that respect. The third strategy involves a conceptual change, for example by
considering a category of parameters as stochastic variation of the observations in-
stead of deterministic parameters. This approach will be used to model atmospheric
signal variability in chapter 4. No strategy alone will be sufficient to resolve the
rank defect. A combination of approaches is expected to be most effective.
The exact formulation of the parameter vector depends on the interferometric con-
figuration, the amount of a priori information, and the goals of the study. In the
remainder of this section we discuss several examples.
If the recovery of topographic information is the goal of the study, interferometric
pairs with a short temporal baseline can exclude the presence of surface deformation
signal, which reduces the rank defect by Another possibility is to use a second
“topographic” interferogram: this will increase the number of observations to
whereas the number of parameters will be increased to as the topographic
height will not change. Similarly, using the same SAR acquisitions in several
interferometric pairs results in a reduction of the atmospheric parameters. Examples
of this approach are presented in section 5.4.9. If the phase unwrapping can be
assumed to be error free, e.g., for high coherence data with smooth topographic
variation and/or small perpendicular baselines, the integer ambiguity number might
be known, which reduces the rank defect by
For deformation studies, topographic information might be available from reference
elevation models or other interferograms. In that case, an expression for the accuracy
of these models needs to be included in the stochastic model, but the topographic
parameter can be eliminated. A considerable improvement is the formulation of the
deformation pattern in terms of a mathematical function. In that case, the number
of unknown deformation parameters can often be reduced considerably.
 3.1 Gauss-Markoff model definition 65

A major source of uncertainty is formed by the atmospheric signal. Note that it is

referred to as signal instead of disturbance, since the effects are spatially correlated
and can be significant, see chapters 4 and 6. Moreover, using different interfero-
metric combinations the atmospheric contribution at the time of some of the SAR
acquisitions can be eliminated. A new application of repeat-pass interferometry,
Interferometric Radar Meteorology, considers the atmospheric signal as the main
parameter of interest. In that case, topography and deformation might be removed
using a reference elevation and deformation model, respectively.
Since the practical definition of the model depends on the goals of the study, we
define a generic model that involves all parameters discussed above.

3.1.3 Generic functional model

A generic linearized functional model for repeat-pass interferometry can be written

as:

where is the part of the design matrix corresponding to observation defined

as:

with The observation vector is the phase

contribution after subtracting the reference body, e.g., spheroid or ellipsoid, with:

with the value for obtained at the reference body. The parameter vector
corresponding to observation is:

The dispersion of the vector of observations can be estimated using empirical coher-
ence measurements or system theoretical considerations, and the diagonal variance-
covariance matrix can be written as:
66 Chapter 3: Functional mode! for radar interferometry

where can be approximated using the estimated coherence values from the coreg-
istered SAR data. As a first order approximation, we assume that there is no cor-
relation between the phase observations in the interferogram, although a resolution
cell strictly has about 50% overlap with the neighboring cells, see section 2.3.8.
In chapter 4 we propose to reformulate the model by lumping the atmospheric slant
delay signal into one parameter transferring this parameter to the
stochastic model, and replace by

where is the variance-covariance matrix of the atmospheric delay parameter

Using this approach, assuming ideal phase unwrapping and a priori knowledge of
either topography or deformation, the design matrix A has full rank.

3.2 Topography estimation

The functional relation between the terrain height above a reference body and
the absolute interferometric phase has been defined in eq. (2.4.12), page 37, to
be

where the “naught” versions and indicate the values for the reference body.
These values change slightly for a point at some elevation. To refine the generic
functional model for the needs of topography estimation it is necessary to reduce
the rank defect in eq. (3.1.6). This could be achieved by eliminating the deformation
component and the two atmospheric components. Regarding deformation, this could
be achieved by choosing a time interval as short as possible, ideally simultaneous
or single-pass acquisitions. For single-pass applications cf. eq. (3.1.7), can be
simplified to

and

For repeat-pass topography estimation the ideal configuration is dependent on the

time interval (which influences temporal decorrelation, surface deformation, vegeta-
tion), atmospheric variability, wavelength, baseline, and the amount and range of
topography within the scene. Rules of thumb regarding atmospheric influence are
given in section 4.7 and in chapter 6.
From fig. 3.2A, the difference between the geometric ranges from point P to satel-
lite 1, and to satellite 2, can be expressed as
 3.2 Topography estimation 67

This range difference can be expressed for each point P on the reference body from
the baseline length B and angle the local look angle and the geometric range

In fig. 3.2A, we find that angle can be written as:

Using the cosine rule this enables an expression for as

which gives an exact but non-linear expression for the range difference:

Since for contemporary repeat-pass interferometric constellations the baseline length

is typically less than 1000 m, the difference between the local look angle is less than
4 arc minutes. This enables an alternative expression for the range difference, the
far field approximation (FFA) (Zebker and Goldstein. 1986). This approximation
considers the local look angles for both sensors to be identical, as in fig. 3.2B. The
parallel look directions provide the approximation for the range difference:

Note that, compared to the exact expression in eq. (3.2.7), the absolute range
has been eliminated. Comparison of the exact and approximated range difference
yields a measure for the approximation error.
68 Chapter 3: Functional model for radar interferometry

Thus, for a given sensor, the approximation error is a function of two variables,
the baseline length B and orientation see fig. 3.3A. Although relatively large
approximation errors are apparent, up to 700 mm for large baselines, it needs to be
stressed that the absolute range difference cannot be directly observed in the
interferogram. This is due to orbit inaccuracies and the unknown phase ambiguity
number. For a specific range rather the minute changes in the look angle are
observed. These changes reflect the deviation between the look angle to a point on
the reference body, and to a point on the topography, cf. fig. 2.10A and 2.10B.
Hence the partial derivatives to the look angle, for the true and the approximated
range differences are

and

The approximation error derived from the difference between eq. (3.2.10a) and
eq. (3.2.10b) is shown in fig. 3.3B. The values represent the range difference for
a 1° change in look angle, which corresponds approximately with the height of Mt.
Everest above the WGS84 ellipsoid. Although the approximation is already reason-
ably good, it can have a non-negligible effect on the sensitivity to height differences,
expressed by the perpendicular baseline Therefore, it is always necessary to
evaluate eq. (3.2.1) iteratively, updating the values for in every iteration. In prac-
tice, however, the iterations are initialized on the neighboring resolution cells, which
makes one iteration sufficient to obtain a value to the cm-level.
 3.3 Deformation mapping 69

3.3 Deformation mapping

As a consequence of the time interval between the subsequent acquisitions, interme-

diate surface deformations will result in a relative range difference, here denoted by
If we write the geometric range for the second acquisition as as a function
of the constellation range difference and the deformation we find

Similar to the sensor constellation, an exact and an approximate expression can be

derived. From fig. 3.4A, can be written as

where is the incidence angle at point P.

If resolution cell P has been subject to a deformation along displacement vector r
in direction this would give a slant range deformation component Using the
cosine rule the new range can be expressed as

with which yields the slant range deformation

In terms of horizontal and vertical deformation components we write

70 Chapter 3: Functional model for radar interferometry

Using the far field approximation a parallel look direction can be assumed. Since
the deformation in practice will be far less then a couple of meters, limited by
a reduction in coherence for large deformations, this approximation can be safely
applied, cf. fig. 3.4B.
A deformation r in direction will result in a slant range deformation component

Thus, only one component of the 3D deformation vector can be observed in the
interferogram. In practice, the combination of ascending and descending orbit inter-
ferograms can be used to find a second component of the deformation vector. Due
to the steep incidence angle of ERS, this satellite is especially sensitive to vertical
deformation.

3.4 Atmospheric mapping

Radar interferometry can be applied to study vertically integrated atmospheric re-

fractivity variations. The parameters and in eq. (3.1.9) reflect every type of
slant-atmospheric phase delay or advancement with respect to the propagation ve-
locity in vacuum (Feynman et al., 1963). This can be due to the velocity variations
along the geometric path between antenna and scatterer or due to the induced bend-
ing of the ray path (Bean and Dutton, 1968). For every point in three dimensional
space, and for every moment in time, the electromagnetic properties concerning ra-
dio wave propagation are governed by the dimensionless refractive index
of the medium. The phase velocity is expressed by
Under normal atmospheric circumstances the refractive index varies only slightly
from its value in a vacuum and therefore often the scaled-up refractivity N
is used, with The slant-atmospheric delay for resolution cell at
acquisition time can be written as (Hanssen, 1998)

where is the incidence angle or zenith angle, H is the height of the satellite
above the position of the scatterer, and is the true slant range. In this equation
we have separated the influence of propagation velocity and the bending of the ray
path. For the propagation velocity part, measured along the true slant path the
incidence angle is a constant. For the bending part, the incidence angle
is a variable along the ray path. Bean and Dutton (1968) have shown that even for
extreme refractivities, the ratio approaches zero for incidence angles
less than 87°. Therefore, for all SAR data available today the slant atmospheric
delay can be considered a function of propagation velocity variations only, which
reduces eq. (3.4.1) to
 3.4 Atmospheric mapping 71

As a rule of thumb, we may use

which is an approximation for the one-way path delay in mm, per km of signal path.
Although eq. (3.4.2) combined with eqs. (3.1.6)–(3.1.9) describe the fundamental
functional relation between the interferometric phase and the integrated refractiv-
ity, decomposition of N in a number of physical parameters can provide new insights
in atmospheric behavior. This is mainly due to the fine resolution of the data, the
high delay accuracy, and the all-weather capabilities. Using only one interferogram,
an ambiguity will be introduced by the coherent superposition of two atmospheric
states This ambiguity restricts the quantitative interpretation of the
atmospheric signal to situations where events are dominant during one of the two
SAR acquisitions. Nevertheless, several studies reported in chapter 6 clearly demon-
strate the feasibility of meteorological interpretation. To circumvent the ambiguity
problem, future applications of the technique can use “cascaded interferograms,” in
which every SAR image appears in two interferograms.

3.4.1 Decomposition of refractivity

For common radar frequencies, the refractivity N can be written as (Smith and
Weintraub, 1953; Kursinski, 1997):

where P is the total atmospheric pressure in hPa, T is the atmospheric temperature

in Kelvin, is the partial pressure of water vapor in hPa, is the electron number
density per cubic meter, is the radar frequency, and W is the liquid water content in
The values and are taken from Smith and
Weintraub (1953), but also results from Thayer (1974) are commonly used. The four
refractivity terms are referred to as hydrostatic term, wet term, ionospheric term,
and liquid term, respectively. Resch (1984) indicated that the first two parts are
accurate to 0.5 percent. Using this representation of the refractivity we decompose
eq. (3.4.2) into

where the four integrals are labeled the hydrostatic delay, the wet delay, the iono-
spheric delay, and the liquid delay, respectively.
As the total signal delay can be up to several meters2 it is evident that the interfer-
ometric data—which were originally wrapped to half the radar wavelength interval
2
The average delay due to a standard troposphere is approximately 2.5 m
72 Chapter 3: Functional model for radar interferometry

(2.8 cm for ERS)—can never capture the full integrated values, unless some sort
of independent absolute calibration is possible. As a result, we can regard every
integral in eq. (3.4.5) as the sum of two contributions. First, a fixed contribution
(bias), which is constant for the whole scene. This value can be subtracted from the
data. Second, a variable contribution dependent of the position in the image. In
terms of the refractivity components, we can write

where is the mean refractivity for all values at height within the interfer-
ogram area, see fig. 3.5. The variable with is the
lateral variation of the refractivity. Consequently, with we could
write eq. (3.4.5) as

As a result of this (relative) representation, it is obvious that the influence of the

four refractivity components depends on their lateral variability within the scene.
We will discuss these contributions in more detail in chapter 6. Vertical variation
of refractivity is of importance if (i) there is topography in the image and if (ii) the
mean vertical refractivity profiles differ between the two acquisitions, see section 4.8.

3.5 Generic model interferogram stacks

Recognition of the rank deficiency in the standard functional model, due to the
superposition of several parameters per observation, has lead to the development of
interferometric methods that combine more than two SAR images. These methods
 3.5 Generic model interferogram stacks 73

eliminate or suppress some parameters at the benefit of others. The three-pass

and four-pass methods to eliminate topography for deformation studies are first
examples of this approach, see Zebker et al. (1994b). Especially the interest in slow
deformation processes and in suppressing atmospheric signal for topography as well
as deformation studies have lead to the use of multiple interferograms, also referred
to as stacking, multi-baseline interferometry, or time-series analysis, see e.g., Ferretti
et al. (1997); Usai and Hanssen (1997); Hanssen and Usai (1997); Sandwell and Price
(1998); Usai and Klees (1999b); Ferretti et al. (1999a).
For multiple interferograms, the generic functional model defined by eq. (3.1.6) for
two-pass interferometry can be extended. When two interferograms are available,
this results in:

where is the part of the design matrix corresponding to observation in inter-

ferogram defined as:

and the parameter vector corresponding to observation and interferogram is:

Here we used the notation and to indicate the atmospheric delay for pixel
at the first SAR acquisitions and the second SAR acquisition In order to
eliminate some of the parameters, both interferograms need to be resampled to the
same (master) grid.
When both interferograms share a common SAR acquisition it is commonly referred
to as three-pass interferometry. Say interferogram I consists of acquisition and
whereas interferogram II consists of acquisition and and surface deformation
doesn’t affect topographic height to the accuracy level we can reorganize
74 Chapter 3: Functional model for radar interferometry

eq. (3.5.1) to:

where and are defined as:

with The parameter vector corresponding to observation in both

interferograms is:

Without a common SAR acquisition the two interferograms are referred to as four-
pass interferometry. As long as the topography can be regarded to be constant, we
can use eq. (3.5.4) with and redefined as:

and the four-pass parameter vector corresponding to observation in both in-

terferograms as:

3.5.1 Three-pass differential interferometry

The concept of three-pass differential interferometry, as demonstrated by Zebker

et al. (1994b), relies on the assumptions that (i) only one of the two interferograms
is affected by surface deformation, and (ii) the deformation does not affect the
topographic height to the accuracy level. Using the flattened phase
we extend eq. (3.5.4) to
 3.5 Generic model interferogram stacks 75

where interferogram I is considered to be the topographic pair (no deformation as-

sumed, absolute phase) and interferogram II is the deformation pair (topography,
deformation, atmosphere, wrapped phase). Resolving for in (3.5.9) and inserting
this in eq. (3.5.10) yields:

where we safely assume that is equal for both interferograms. With

(the baseline ratio) we reformulate the model of observation equations, cf. eq. (3.5.1),
to a mixed model of form (Teunissen, 2000):

assuming A and B have full rank.

where is the differential design matrix corresponding to observation defined

by:

and the parameter vector corresponding to resolution cell is:

The three center elements of are associated with the three atmospheric param-
eters. From eq. (3.5.15) it is clear that the baseline scaling operation affects the
atmospheric signal in the differential interferogram. In general, a large absolute
baseline ratio needs to be avoided as it amplifies the atmospheric signal in one of
the two images contributing to the topographic pair. This is equivalent to choosing a
topographic pair with a large baseline and a deformation pair with a relatively small
baseline. In the limit, a zero baseline for the deformation pair would make the topo-
graphic pair unnecessary and has normal (non-scaled) atmospheric circumstances.
The considerations for choosing the topographic pair are generally
76 Chapter 3: Functional model for radar interferometry

1. to obtain at least the same height accuracy as the deformation

pair,
2. to obtain coherent data, and
3. terrain circumstances such as relief, vegetation, and temporal decorrelation.

Table 3.2 and fig. 3.6 show the consequence of the baseline scaling ratio for the
amplification or reduction of the atmospheric signals in the common image and
in the slave of the topographic pair. A reduction of the atmospheric signal in
both images can only be obtained for all other baseline ratios result in
the amplification of atmospheric signal in either one or both of the images of the
topographic pair.
Regarding the atmospheric signal as stochastic—as will be proposed in chapter 4—
implies that the baseline ratio must be used to weight the variance and covariance
of the atmospheric signal . In spectral approximations, the spectra are multiplied
by and (Bracewell, 1986). The variance for a zero-mean signal will be
amplified by and

3.5.2 Four-pass differential interferometry

In some cases no suitable SAR acquisition is available to form the topographic pair
together with one acquisition from the deformation pair. However, there might be
 3.5 Generic model interferogram stacks 77

another suitable interferometric combination that could serve as topographic pair.

This situation occurs, e.g., when the perpendicular baselines for any combination
with one of the acquisitions of the deformation pair are too large, or have a temporal
baseline being too long. In those cases the two independent interferometric pairs
can be used to apply four-pass differential interferometry.
A problem related to this technique is that the alignment procedures might fail. The
correlation between the speckle pattern generally diminishes with increasing baseline.
Nevertheless, there might be prominent reflectors that appear in both images. Since
no coherent phase differences are necessary between both pairs, alignment criteria
are less stringent. The most important factor for these applications is the roughness
of the topography.
Assuming the topographic pair, labeled I, is flattened, unwrapped, and does not
contain any deformation, the functional relations in eqs. (3.5.9) and (3.5.10) become

Following the same procedure as for three-pass differential interferometry, we end

up with the mixed model of eq. (3.5.14), with defined as

and the parameter vector corresponding to resolution cell as

Using table 3.2 the weighting of the atmospheric signal in the topographic pair can
be determined. It is clear that a topographic pair with a long baseline is to be
preferred.

3.5.3 Linear integer combinations

Linear combinations of interferograms as in eq. (3.5.14) can be used to decrease rank

deficiency. However, an important limitation of scaling one of the interferograms
with baseline ratio is that this interferogram needs to be unwrapped before
scaling, introducing errors that propagate into the final results. The necessity for
the phase unwrapping is nevertheless limited to In these situations,
scaling a wrapped interferogram will introduce fractional, hence erroneous, phase
jumps in the data. Massonnet et al. (1996c) showed that this condition does not
apply for hence linear integer combinations. In these cases, the original
fringes remain at their position but new fringes may be introduced between them.
Since an integer scaled interferogram corresponds with an integer scaled perpen-
dicular baseline, with it is possible to form the difference with
78 Chapter 3: Functional model for radar interferometry

an interferogram with perpendicular baseline The resulting difference interfero-

gram will have a perpendicular baseline of By choosing factor a so
that is very small, the difference interferogram will have much less fringes, and
might be easier to unwrap. Although the accuracy of this unwrapped (first-order)
difference interferogram is not optimal—the small baseline results in a large height
ambiguity—it can be subtracted from the original interferograms. This yields an in-
terferogram with residual fringes, which might be more successfully unwrapped than
the original one. Adding the residual unwrapped fringes to the unwrapped fringes of
the linear combination consequently yields the final results in which possible errors
in the linear integer combination are eliminated. Figure 3.7 shows an example of
this procedure for an area in Cornwall, United Kingdom. It is important to note
that the integer scaling of one of the baselines amplifies noise in that interferogram
as well. In practice, the scaling factor proves to be limited to
It might be elucidating to express the generic functional model in terms of topo-
graphic height only, assuming no deformation or atmospheric signal. In that case
we can write the functional relation between height and the flattened, wrapped
interferometric phase in interferogram I and II as

which can be simplified when denoting the height ambiguity for the two interfero-
grams by and

with and The linear integer combination

with can be written as

with

Therefore, by choosing factors and such that is large, phase gradients cor-
respond with large height gradients. As a result, at the expense of increased phase
noise, the height gradients can be such that the likelihood for estimating the correct
integer value for is increasing considerably. Therefore, this methodology can be
used to reduce the solution space in a way comparable to the two carrier frequencies
(L1,L2) approach in GPS ambiguity fixing.
3.5 Generic model interferogram stacks 79
 Chapter 4
Stochastic model for radar interferometry

This chapter presents the stochastic part of the Gauss-Markoff model for radar
interferometry. It revisits the necessary theoretical concepts before focusing on
the single-point statistics and multiple-point statistics. The quality of an inter-
ferogram is described in terms of the first and second moments. Error sources
introduced by scattering, geometry, instrument, orbit, atmosphere, and signal
processing are discussed. Models are proposed to describe atmospheric signal
due to turbulent mixing and vertical stratification. Error propagation is used
to derive a generic stochastic model.
key words: Stochastic model, Quality analysis, Error sources, Error propaga-
tion, Atmosphere

In the previous chapter, the functional part of the mathematical model has been de-
fined. The functional model relates the interferometric observations to the unknown
parameters of interest, such as topography, deformation, or atmospheric signal. In
this chapter we will focus on the stochastic part of the mathematical model, which
describes the dispersion of the observations in terms of variances and covariances.
A key aspect in the definition of this model is the analysis of the error sources and
the way these errors propagate into the stochastic model. As the sub-title of this
book “Data Interpretation and Error Analysis” suggests, the analysis, propagation,
and mathematical formulation of errors is indissoluble from a correct interpretation
of the data.
In this chapter, we will derive and discuss the most important error sources in repeat-
pass radar interferometry, starting with a concise theoretical background of the used
concepts in section 4.1. Section 4.2 covers the single point observation statistics
in the absence of biases or correlations between resolution cells. The concept of
coherence and its relation with the signal-to-noise ratio is discussed in section 4.3,
followed by a systematic treatment of the error sources affecting coherence and single
point statistics. This evaluation includes instrument noise, processing induced errors,
and temporal decorrelation. Sections 4.5–4.8 describe errors affecting the correlation
between different resolution cells in the interferogram, starting with integer phase
ambiguities, followed by orbit induced errors in 4.6 and atmospheric induced errors
in section 4.7 and 4.8. In section 4.9 all error sources are combined in a joint
stochastic model, and the propagation of phase errors into heights, deformations or
82 Chapter 4: Stochastic model for radar interferometry

atmospheric parameters is derived.

4.1 Theory

In this introduction a brief review is presented of the main statistical parameters

used to describe the behavior of radar interferometric data.

4.1.1 First-order moment or expectation value

The random variable at location in an image has an expectation

value

where is the probability density function of The expectation value

can be a function of the location

4.1.2 Second-order moments

The dispersion of the observations about their expectation value is described by the
second-order moments. Three second-order moments are the variance, the covari-
ance and the structure function or variogram. Note that structure function is a term
frequently used in turbulence literature, while variogram originates from geostatis-
tics. Covariance and structure function are often referred to as the two structural
functions.
The variance is defined as (Box et al., 1994)

and the covariance as

which is a function of the locations and

The third second-order moment is the structure function, which is defined as the
variance of the increment

where is the structure function or variogram and is the semi-

variogram (Journel and Huijbregts, 1978).
Autocorrelation
The autocorrelation function is a standardized version of the covariance function,
bounded to the interval [–1,1], with
 4.1 Theory 83

or

for a (second-order) stationary process. In contrast to the covariance function, the

autocorrelation function is independent of the magnitude of the data values.

4.1.3 Positive definiteness

A covariance function, empirical or analytical, is positive (semi)-definite whenever

it fulfills the following three conditions (Journel and Huijbregts, 1978):

this is always true for and

Schwarz’s inequality.

Often the covariance function will reach zero values for a distance larger than a
specific value, indicating that there is no correlation for two values which are further
apart than that distance. This distance is referred to as the range of the covariance
function. This implies with eq. (4.1.12) that the value of the structure function after
that range will be or which is called the sill value.
If the covariance function is positive semi-definite, the variance of any linear com-
bination of variables must be positive (Wackernagel, 1995). Therefore, the
covariance matrix C is positive semi-definite if and only if the covariance function
fulfills these three requirements.

4.1.4 Stationarity

To apply specific statistical measures it is necessary to determine whether the un-

derlying physical process can be considered statistically stationary. Unfortunately,
unless infinitely many sample functions are available of this process, which reach to
infinity, it is never possible to answer this question (Newland, 1993). Therefore, it
is often tried to apply the mathematics under certain assumptions of the physical
behavior of the process, or to relax the stationarity conditions. In terms of nested
sets, we define the set of strictly stationary random functions as a subset of the set
of second-order stationary random functions, which is a subset of intrinsic random
functions, see fig. 4.1. All these sets are contained in the general set of random
functions (Journel and Huijbregts, 1978) (Chilès and Delfiner, 1999).
Proceeding from the large sets to the smaller subsets we find:

Intrinsic random functions. A random function is intrinsic when (i) the

expectation value is constant and exists for every location :
84 Chapter 4: Stochastic model for radar interferometry

and (ii) for all vectors the increment has a

constant finite variance:

Therefore, this set of functions is also referred to as random functions with

stationary increments. This does not imply that the variance and covariance of
can be uniquely determined. Written in matrix notation the expectation
of the increment and its variance are

From (4.1.10) it follows that there are more combinations of variances and
covariances of which result in the same value for Note that
the definition of the structure function does not require the existence
of a constant mean and finite variance of (Goovaerts, 1997).
Second-order stationarity.
A random function is stationary of order 2 when the first two moments
are stationary. In that case, the two characteristics of the intrinsic hypothesis
hold, and (iii) for each pair the covariance exists and is only
dependent on the distance r:

The existence of the covariance implies that the right-hand side of eq. (4.1.10)
is known. Therefore, it is always possible for a second-order stationary function
to derive the structure function from the covariance function

Therefore, under the hypothesis of second-order stationarity, the covariance

and the structure function are in fact equivalent (Journel and Huijbregts,
1978). Note that there are physical phenomena, such as Brownian motion,
which have an infinite dispersion. For such phenomena does not exist,
and although they might fulfill the intrinsic hypothesis they cannot be consid-
ered second-order stationary. On the other hand, since practical measurements
span a finite range, second-order stationarity can be assumed within this range.
This corresponds to an hypothesis of quasi-stationarity.
Strict stationarity. A random function is strictly stationary if the whole
probability density function (i.e., all statistical moments) is invariant under
translation. Note that strict stationarity is a severe requirement, and usually
the simple term stationarity refers to second-order stationarity.
 4.1 Theory 85

Often the covariance or structure function are only used in a limited region. By
limiting the hypothesis of second-order stationarity to this region, the function can
be considered quasi-stationary.
It is important to note that stationarity is a property of the random function or
process, not the data, due to the finite and discrete nature of the data. Therefore,
in practice, stationarity is a decision made by the user, not a hypothesis that can be
proved or rejected by data. Often it is even only a characteristic of the used model
and not of the physical properties under study (Goovaerts, 1997).
Homogeneity and isotropy
Homogeneity is the equivalent of second-order stationarity and is used especially for
2D and 3D random fields (Balakrishnan, 1995). A second-order stationary random
field is isotropic if it is homogeneous and if the covariance function only depends on
the length of the increment and not on its orientation (Chilès and Delfiner. 1999),
(Balakrishnan, 1995):

Ergodicity
For an ergodic signal, the spatial (or temporal) and ensemble averages are equivalent.
For example, interferometric coherence estimation using a spatial window relies on
the assumption of ergodicity.

4.1.5 Scale-invariance and fractal dimension

For a scale-invariant (scaling) signal there is no characteristic scale in the process,

so that power laws prevail over a large range of scales. For a scaling process
the power spectrum exhibits a power-law behavior:
86 Chapter 4: Stochastic model for radar interferometry

for all discrete wavenumbers within the scaling range. Note that the expression
power spectrum is also commonly referred to as wavenumber spectrum, energy spec-
trum, variance spectrum, periodogram, and spectral density. The spectral exponent
contains information on the degree of stationarity of the signal (Mandelbrot, 1983).
If the signal is stationary, i.e. is statistically invariant for translation.
For the signal is non-stationary with stationary increments, so the struc-
ture function will be stationary (Ishimaru, 1978; Agnew, 1992; Davis et al., 1994).
Many geophysical processes belong to the latter category over a bounded range of
scales, which restrict the process to its physically accessible values.
Fractal dimension
Scale-invariant processes can be conveniently described using power-law statistics or
fractal statistics. Since many processes in nature are scale-invariant within a certain
range, the fractal dimension D is an elegant parameterization of the behavior of a
process, equivalent to the spectral exponent A fractal that is statistically self-
similar exhibits the same behavior in all dimensions, and is by definition isotropic.
For example, the isoline curves (horizontal planar sections) of a topography map
are self-similar. The fractal surfaces that are important in this study are self-affine
fractals. For such fractals the scale in the vertical plane differs from the scale in the
horizontal plane (Chilès and Delfiner, 1999).
For a one-dimensional signal or two-dimensional random field the relationship be-
tween the spectral exponent Haussdorff measure H, and fractal dimension or
is (Turcotte, 1997):

For a self-affine fractal we find This category

of processes is referred to as fractional Brownian motion (fBm). Standard fBm or
random walk is a subset within this category with H = 0.5, or
(Mandelbrot and Ness, 1968).
The fractal dimension or power-law exponent is a measure for the smoothness of a
signal or surface. Figure 4.2 shows three characteristic fractal simulations: “white
noise,” “1/ f-noise” (where f is the continuous wavenumber), and “standard fBm”
in a one-dimensional realization (left column), two-dimensional realization (middle
column), and the power spectral shape in a log-log plot (right column). Notice that
a lower fractal dimension or steeper slope of the power spectrum (higher value for
corresponds with a smoother surface.
The three examples in fig. 4.2 demonstrate the relationship between stationarity
and the spectral exponent or fractal dimension. White noise is strictly
stationary: and For
the signal is not strictly stationary: taking for convenience, the variance
is proportional to even though Since the structure
function is independent of position , the signal
has stationary increments and is intrinsic. Nevertheless, for a limited range such as
depicted here, the signal can be considered second-order stationary.
 4.2 Single-point observation statistics 87

The concepts of scale-invariance will be used in section 4.7 to describe and model
the general behavior of atmospheric signal in the SAR interferograms.

4.2 Single-point observation statistics

Radar interferometry can be regarded as a parameter estimation problem, where

the observations consist of complex-valued numbers for all resolution cells and the
parameters depend on the goal of the study, as discussed in chapter 3. In that
chapter, the standard Gauss-Markoff model, defined in eq. (3.1.1), was proposed
to obtain optimal estimates of the parameters and their corresponding covariance
matrix, see eqs. (3.1.3) and (3.1.4), respectively. The functional relation between
parameters and observations has been defined. Nevertheless, correct assessment
of the dispersion of the observations in the stochastic model is the key issue for
evaluating adjustment procedures and assessing the final data quality. Let us look
at this dispersion in more detail.
Many authors have evaluated the (complex) variance of single-look or multilook
interferometric observations (Li and Goldstein, 1990; Tough et al., 1995; Joughin,
1995; Touzi et al., 1999; Lee et al, 1994; Just and Bamler, 1994; Bamler and Just,
1993; Davenport and Root, 1987; Rodriguez and Martin, 1992). In these approaches,
however, the covariance between observations was considered to be restricted to
neighboring resolution cells. This hypothesis was mainly supported by the fact
that the bandwidth of the SAR in range and azimuth direction is smaller that
the frequency range determined by the Nyquist frequency. This effect results in
resolution cells that are slightly larger than the posting of the observations, and is
accounted for, e.g., by taking an effective number of looks instead of the nominal
amount of looks. Correlations between resolution cells that are not neighboring
are generally not accounted for. Ignoring (in first approximation) the correlations
between neighboring resolution cells results in a diagonal variance-covariance matrix,
cf. eq. (3.1.10):

where can be approximated using the estimated coherence values from the aligned
SAR data. This derivation will be discussed in section 4.2.2.
There are two reasons why this approach is generally insufficient for parameter
estimation from interferometric radar data. First, the interferometric observation
is inherently relative. This implies that, instead of observing single (multilooked)
resolution cells, one is generally interested in the difference between two spatially
separated resolution cells, as implicitly visualized in an interferogram. This intro-
duces the need to analyze covariances between the data. Secondly, for all repeat-pass
interferometric configurations the influence of spatially varying signal propagation
effects cannot be ignored. As we will see in section 4.7, atmospheric signal generally
88 Chapter 4: Stochastic model for radar interferometry

obeys a power-law behavior. This implies that a larger separation between resolu-
tion cells results in a larger variance of the difference. Hence, the varying correlation
between resolution cells, dependent on their spatial separation, cannot be ignored,
and the variance-covariance matrix in eq. (4.2.1) will become a full matrix, as:

where is the variance-covariance matrix of the atmospheric delay parameter

defined by lumping the atmospheric slant delay signals and during the two
acquisitions into one parameter Of course, it depends on the pa-
rameter of interest (topography, deformation, or atmosphere) whether this approach
should be followed. If the main objective of the analysis of the interferometric data
is atmosphere studies, it is obvious that the atmospheric delay parameter should be
included in the functional model, not in the stochastic model. As a consequence,
for atmospheric studies the absolute variance of the data is much better than for
topography or deformation studies.
In the following subsections we summarize the analysis of SAR and InSAR data
statistics for single (multilooked) resolution cells.
 4.2 Single-point observation statistics 89

4.2.1 Complex PDF SAR resolution cell

The response of a resolution cell on earth to the arriving radar pulse is strongly
dependent on the scattering mechanisms involved. Although this is an extended
field of study, only two mechanisms will be discussed here: distributed scattering
(also referred to as Gaussian or Rayleigh scattering) and point scattering (Madsen,
1986; Bamler and Hartl, 1998).
For SAR systems where the size of a resolution cell is many times larger than the
radar wavelength, many terrain elements or scatterers contribute to the response
for one resolution cell (Krul, 1982; Curlander and McDonough, 1991). In practice
it is therefore not possible to determine the response of individual scatterers within
a resolution cell. This results in the well known, highly unpredictable speckle char-
acteristics of SAR magnitude images (Goodman, 1975). The measured reflection is
rather written as a sum of many (sub) scatterers, see fig. 4.3. Using a defined set of
assumptions, it is possible to apply the central limit theorem, which defines the ob-
servations as complex (circular) Gaussian random variables. These assumptions
are (Madsen, 1986; Bamler and Hartl, 1998):

1. No single scatterer should dominate the others in a resolution cell. This as-
sumption generally holds for, e.g., agricultural fields, forests, deserts, and many
other natural scatterers.

2. The phase of every individual scatterer must be uniformly distributed between

and When the slant-range resolution is much larger than the radar
wavelength, a very large range of phase shifts is apparent. Folded into the
interval this leads to a uniform distribution.

3. The phases of the individual scatterers must be uncorrelated. This is reason-

able as correlated scatterers are gathered into several (uncorrelated) scatter
centers.
90 Chapter 4: Stochastic model for radar interferometry

4. The amplitude and the phase of every scatterer must be uncorrelated. This
assumption holds because the phase delay due to the propagation of the signal
is independent of the scattering magnitude.

The probability density function of a complex circular Gaussian variable y (or the
joint PDF of its real and imaginary component) is written as (Dainty, 1975; Dav-
enport and Root, 1987):

where

In the context of radar observations, the factor in (4.2.3) is often substituted by

(Bamler and Hartl, 1998)

where is the intensity (or power) of the resolution cell and is its
expectation value. From equation (4.2.3) it is clear that the real and imaginary
parts of a complex circular Gaussian variable are uncorrelated. From (4.2.3) and

we find the Jacobian and derive the joint probability density function for
the amplitude a and phase random variables:

The marginal PDF of a is obtained integrating out, between and This

yields

system theoretical description of was given by Madsen (1986),

where N is the number of independent scatterers within the resolution cell, is the complex
backscatter coefficient per scatterer, and h is the coherent system impulse response function. Similar
expressions are discussed by Bamler and Schättler (1993); Bamler and Hartl (1998).
 4.2 Single-point observation statistics 91

Equation (4.2.9) is the Rayleigh distribution (Papoulis, 1968), see fig. 4.4A. The
marginal PDF of is found integrating (4.2.8) over a, between 0 and

Equation (4.2.10) describes a uniform distribution. From (4.2.8), (4.2.9), and (4.2.10)
it follows that a and are uncorrelated, since

The pixel intensity variation for a distributed scene, expressed by the exponential
PDF, is known as speckle. Assuming ergodicity, averaging of resolution cells is often
applied to reduce the effect of speckle. As a result, the PDF of the intensity value
of N averaged resolution cells can be described by a PDF with 2N
degrees of freedom (Raney, 1998):

Note that for N = 1 this equals the exponential PDF, while for N it equals a
Gaussian PDF.
In table 4.1 the probability density functions of the intensity p, magnitude a, and
phase of a resolution cell are summarized (Madsen, 1986; Bamler and Hartl, 1998),
(Raney, 1998):

4.2.2 Complex PDF interferometric resolution cell

In section 4.2.1 the behavior of a single SAR resolution cell, under the assumption
of a distributed scattering mechanism, was described by circular Gaussian statistics,
see eq. (4.2.3). Although the PDF of the phase has a uniform distribution, see
table 4.1, the phase of the complex product of two circular Gaussian signals is not
necessarily uniform, as long as the two signals have some degree of correlation.
92 Chapter 4: Stochastic model for radar interferometry

The joint PDF of two circular Gaussian signals and with

and can be written as (Goodman, 1963; Hannan and Thomson,
1971; Lee et al., 1994; Just and Bamler, 1994; Tough et al., 1995; Bamler and Hartl,
1998)

where is the complex covariance matrix, defined by

and is its determinant (Tough et al., 1995):

The complex coherence (or complex correlation coefficient) is defined as

As the interferometric combination of and can be written as the Hermitian

product:

we find that is equal to the numerator of the complex coherence. Since the
denominator is real-valued, the phase of the complex correlation is the expected
phase of the interferogram. The magnitude with is a measure of
the phase noise. Applying a multilook procedure with N looks, during this
multiplication, eq. (4.2.17) can be written as
 4.2 Single-point observation statistics 93

Rodriguez and Martin (1992) showed that the maximum likelihood estimator of the
interferometric phase, for distributed scattering mechanisms, is

which is unbiased modulo

The joint probability density function of the interferometric amplitude a and phase
for L independent looks (L < N for SAR systems with a limited bandwidth) can
be shown to be (Goodman, 1963; Barber, 1993b; Lee et al., 1994; Tough et al., 1995):

where is the average power of the two image segments.

is the modified Bessel function of the third kind (also referred to as Hankel’s function
(Gradshteyn et al., 1994)), and is the complex correlation coefficient. The gamma
function is defined as:

or as

The marginal probability density function for the interferometric phase is obtained
by integrating over all amplitudes a (Tough et al., 1995):

with Note that the finite summation in (4.2.23) disappears for

the single look case (L — 1).
Barber (1993a), Lee et al. (1994), and Joughin and Winebrenner (1994) found an
equivalent expression, using a hypergeometric function:
94 Chapter 4: Stochastic model for radar interferometry

where the hypergeometric function is defined as (Oberhettinger, 1970; Gradshteyn

et al., 1994):

For single-look data, L = 1, eqs. (4.2.23) or (4.2.24) reduce to (Just and Bamler,
1994; Tough et al., 1995):

Figure 4.5 shows the shape of the probability density functions for different coherence
levels. Figures 4.5A, B, and C correspond with multilook levels of 1, 10, and 20,
respectively.
With the PDF’s defined in eqs. (4.2.23) and (4.2.24), the phase variance resulting
from can be easily derived using eq. (4.1.2):

For single-look data the phase variance can be expressed in a closed form using (Bam-
ler and Hartl, 1998)
 4.2 Single-point observation statistics 95

where is Euler’s dilogarithm, defined as (Abramowitz and Stegun, 1970)

and

For numerical evaluations, eq. (4.2.23) appears to be faster and more accurate than
eq. (4.2.24), since it does not involve an infinite summation. In fig. 4.6, the standard
deviation of the interferometric phase is evaluated as a function of the coherence
level, for three different multilook levels.

Statistics for point scatterers For point scatterers, which can be regarded as
a constant, deterministic signal plus noise, the variance according to eq. (4.2.27)
is overestimated. For those cases, where is close to 1, the variance of the phase
is given by (Bendat and Piersol, 1986), (Just and Bamler, 1994):

which follows from the Cramér-rao bound for the phase variance, as derived by
Rodriguez and Martin (1992):

In fig. 4.6 the behavior of eq. (4.2.31) is indicated by the short bold line between
0.9 < 1. If a stack of coregistered interferograms is available, a pixel with a
96 Chapter 4: Stochastic model for radar interferometry

relatively large amplitude in every interferogram is an indication of possibly stable

phase behavior. This can be a first criterion for the selection of permanent scatterers,
see Ferretti et al. (2000). Simply said, for point scatterers the amplitude behavior
is an indication for the phase variance.

4.3 Coherence and SNR

The complex coherence between two zero-mean circular Gaussian variables and
is defined as (Born et al., 1959; Foster and Guinzy, 1967; Papoulis, 1991)

In this definition a coherence value can be assigned to every pixel in the interfero-
gram. The coherence can be used as a measure for the accuracy of the interferometric
phase, as discussed in section 4.2.2, or as a tool for image classification. Ideally, the
expectation values in eq. (4.3.1) are obtained using a suite of observations for ev-
ery single pixel, i.e., a large number of interferograms acquired simultaneously and
under exactly the same circumstances. In this case, an ensemble average could be
used to obtain the expectation values in eq. (4.3.1). Unfortunately, this procedure
is not feasible, as every full-resolution pixel is observed only once during each SAR
acquisition.
In practical situations the accuracy of phase observations of a uniform region is
assumed to be stationary. Under the assumption of ergodicity it is possible to
exchange the ensemble averages with spatial averages, obtained over a limited area
surrounding the pixel of interest. This assumption is used to obtain the maximum
likelihood estimator of the coherence magnitude over an estimation window of N
pixels (Seymour and Cumming, 1994),

Since (for simplicity often denoted as is used in many of the equations defined
in the previous section, it is important to consider its statistics.
Touzi and Lopes (1996) have demonstrated that the probability density function of
the coherence magnitude estimator can be expressed as a function of the absolute
value of the true coherence and the number of independent samples, L > 2 as

This equation holds for distributed scattering mechanisms. The expectation of

is derived as
 4.3 Coherence and SNR 97

which is plotted in fig. 4.7A.

It appears that the estimate is biased towards higher values, in particular for
low coherence and/or small estimation windows (Joughin and Winebrenner, 1994;
Tough et al., 1995; Touzi et al., 1996, 1999). The variance of the coherence magnitude
estimator is derived from eq. (4.3.3) as well:

which is plotted in fig. 4.7B. The Cramér-Rao bound, which is a lower bound for
the variance (assuming unbiased estimation) is defined as

and indicated as a reference by the dashed line in fig. 4.7B. From figs. 4.7A and 4.7B
it can be observed that the Cramér-Rao bound is a good approximation as long as the
estimates are unbiased (Priestley, 1981). Removal of the bias in by inverting
eq. (4.3.4) is only possible when is low or when L is sufficiently large (Touzi
et al., 1999). For biased estimates, eq. (4.3.5) is to be preferred over eq. (4.3.6).

Phase corrected coherence The coherence as determined using eq. (4.3.2) is

affected by interferometric phase variation due to (i) noise in the data and/or to
(ii) systematic phase variation within the coherence estimation window. Systematic
phase variation is a result of effective path length variations over the scene, which
can be attributed to topographic, atmospheric, or deformation gradients. If the
98 Chapter 4: Stochastic model for radar interferometry

coherence estimation is solely used for deriving the phase statistics, eq. (4.3.2) is
correct. However, if the coherence is used for, e.g., classification of terrain types,
derived from the noise level of the data, the systematic phase variation needs to
be excluded. This necessitates an adapted definition of the coherence, the “phase-
corrected” coherence defined as (Hagberg et al., 1995: Monti Guarnieri and
Prati, 1997; Dammert, 1997)

where is the systematic phase component for each pixel.

Relation with SNR The absolute value of the coherence which ranges be-
tween 0 and 1, is equivalently described as a function of the signal to noise ratio
(SNR) (Foster and Guinzy, 1967; Prati and Rocca, 1992; Zebker and Villasenor,
1992; Bamler and Just, 1993):

4.4 Sources of decorrelation

The definition of decorrelation can be rather arbitrary. Here we define decorrelation

as the noise caused by error sources that have a correlation length smaller than a
regular coherence estimation window. This implies that orbit errors and errors due
to atmospheric heterogeneities will be discussed separately.
In the previous section we have shown that phase noise in interferograms can be ex-
pressed as a function of the absolute value of the complex coherence. The complex
coherence can either be estimated from the interferometric data, based on ergod-
icity assumptions, or derived theoretically based on the known sensor characteris-
tics, acquisition circumstances, and signal processing algorithms. Several sources of
decorrelation, expressed by their correlation terms, can be distinguished, such as

baseline or geometric (surface) decorrelation caused be the difference

in the incidence angles between the two acquisitions;
Doppler centroid (surface) decorrelation caused by the differences in the
Doppler centroids between the two acquisitions;
volume decorrelation caused by penetration of the radar wave in the
scattering medium;
thermal or system noise caused by the characteristics of the system,
including gain factors and antenna characteristics;
temporal terrain decorrelation caused by physical changes in the
terrain, affecting the scattering characteristics of the surface, and
processing induced decorrelation which results from the chosen
algorithms, e.g., for coregistration and interpolation.
 4.4 Sources of decorrelation 99

It can be shown that the listed correlation terms are multiplicative (Zebker and
Villasenor, 1992), which results in a total correlation or coherence:

In the following sections these correlation factors will be discussed in more detail.

4.4.1 Thermal decorrelation

The influence of thermal noise on the interferometric phase can be derived theoret-
ically by determining the signal-to-noise ratio of a specific system (Zebker et al.,
1994c; Zebker, 1996). In table 4.2 the SNR for ERS-1 and ERS-2 is derived.
The average received signal power can be determined using the radar equation,
which can be written in the following form (Ulaby et al., 1982; Zebker, 1996):

where
the transmit power [Watts],
the distance between the antenna and the resolution cell [m]

the altitude of the satellite [m],

the incidence angle,
the directivity gain [dimensionless],
the antenna area
the radar wavelength [m],
the scattering area and
the normalized radar cross section [dimensionless].
The scattering area is defined as the size of the area which is illuminated by
a single pulse. In azimuth direction, this area is determined by the width of the
antenna beam multiplied with the average range between the antenna and the
ground R:

where is the length of the antenna. In range direction, the range of a single pulse
is determined by pulse length and the speed of light c, followed by a projection
from slant-range to ground-range:

The normalized radar cross section is the fraction of the power intercepted by the scatterer
and the power returned to the radar. The parameters that affect are primarily surface geometry
and moisture content. The surface geometry is determined by the roughness, slope, and vertical or
horizontal heterogeneity. The electrical properties of soil and vegetation, expressed by the dielectric
constant are strongly influenced by the moisture content.
100 Chapter 4: Stochastic model for radar interferometry

With eqs. (4.4.3) and (4.4.4) we find

The power of the thermal noise in the receiver system is (Curlander and Mc-
Donough, 1991; Zebker et al., 1992; Zebker, 1996)

with

k, Boltzmann’s constant [J/K],

the receiver noise temperature
the system bandwidth

The signal-to-noise ratio is found dividing the signal power by the power of the noise,

and is calculated for ERS-1/2 in table 4.2. It shows that the SNR value is influenced
by two classes of parameters, those which are determined by the design of the radar
system and those determined by the scene. The latter is only expressed by
Therefore, we can determine the signal-to-noise ratio as a function of see fig. 4.8.
The correlation coefficient between two complex signals and
consisting of a common part c and two thermal noise parts and can be written
as (Foster and Guinzy, 1967; Zebker and Villasenor, 1992):

The two noise components are assumed to be uncorrelated, and the signal is uncor-
related with the noise, we can write

assuming Using the definition of the signal-to-noise ratio

we can obtain (Zebker and Villasenor, 1992)

3
The temperature T is determined by matching the product k T with the correct noise power
spectral density (Curlander and McDonough, 1991).
 4.4 Sources of decorrelation 101

if the SNR values are different for the two sensors, as with ERS-1 and ERS-2.
Thus, combining eqs. (4.4.7) and (4.4.10) we found a direct relationship between
the radar system parameters and the thermal correlation coefficient. Note that this
relationship cannot be regarded separately from the scene’s radar cross section
102 Chapter 4: Stochastic model for radar interferometry

4.4.2 Geometric decorrelation

Geometric (baseline) decorrelation is a result of a difference in the angle of incidence

between the two sensors at the earth’s surface (Gatelli et al., 1994). The main
concepts of geometric decorrelation have been described in section 2.5.5. see page 47.
There we found that the amount of decorrelation increases linearly with the amount
of spectral shift between the two SAR acquisitions, assuming a rectangular spectral
window, see fig. 2.14. Although the amount of spectral shift, i.e., the fringe frequency,
can be due to deformation or atmosphere as well as topography, we focus here on
the latter.
The critical baseline is the baseline causing a spectral shift equal to the
bandwidth It is a function of the wavelength the incidence angle and
the topographic slope defined positive away from the satellite:

The geometric decorrelation can now be simply defined as

and is plotted for three topographic slopes in fig. 4.4.2. For a flat horizontal terrain,
the ERS critical baseline is approximately 1.1 km. A rectangular spectrum was
assumed in this derivation.
A priori filtering can eliminate the geometric decorrelation to a considerable extent,
at the expense of spatial resolution. Slope-adaptive spectral shift filtering needs an
estimate of the topographic slopes, either from an a priori DEM or from the local
fringe frequency. It is obvious that fringe frequency estimation needs a window of a
considerable size to operate, which will limit the filtering for very rugged terrain.
 4.4 Sources of decorrelation 103

4.4.3 Doppler centroid decorrelation

The azimuth equivalent of geometric decorrelation is caused by the difference in

Doppler centroid frequencies, between both acquisitions, as discussed in sec-
tion 2.5.5. Assuming uniform spectral weighting, the coherence factor decreas-
ing linearly with increasing as in

where is the bandwidth in azimuth direction. For the yaw-steered SAR onboard
ERS, this effect is minimal. However, for interferograms between ERS-1 and ERS-2
acquisitions, as well as ERS-2 interferograms with images acquired after 7 February
the effect can be significant (Swart, 2000). If the final interferometric combi-
nation is known, it is common to use the average Doppler of the two images during
the SAR processing. In that case, azimuth filtering is not strictly necessary. How-
ever, if pre-processed SLC images are used, or if interferograms need to be stacked
and Doppler centroids cover a wide range of values, azimuth filtering can improve
the interferogram quality significantly.

4.4.4 Temporal and volume decorrelation

Temporal decorrelation occurs when the distribution of wavelength-scale scatterers

within a resolution cell, or their electrical characteristics, differs as a function of
time between the first and the second data acquisition, uncorrelated with other
resolution cells. The measured interferometric phase is the difference between the
phase values of corresponding aligned resolution cells during the acquisitions, hence

4
At 7 February 2000, an error in two out of three gyroscopes of ERS-2 reduced its Doppler
steering capabilities.
104 Chapter 4: Stochastic model for radar interferometry

Although considerably simplified, it is helpful to consider both initial phase observa-

tions consisting of two deterministic components and a stochastic or noise component

where is the deterministic phase component due to the geometry of the

sensor and the targets on earth at acquisition time In this expression, signal
delay contributions are also considered in the geometric component of the phase.
The deterministic phase contribution of all scatterers within the resolution cell is
denoted by and is additional noise. The scattering component of the
phase is a function of the distribution of scatterers and the viewing direction of the
radar. It is a deterministic value, since a repeated observation under exactly the same
circumstances will yield the same phase value. Temporal decorrelation effectively
changes the contribution of for different Useful coherent interferometric
phase observations are only possible whenever the scattering components
are more or less the same during each of the two acquisitions, expressed by the
temporal coherence cf. eq. (4.4.1). High temporal coherence values are
generally obtained for areas without lush vegetation (i.e., arid areas or polar areas).
Especially for vegetated areas longer wavelength radar data, for instance L-band,
are preferred over C-band or X-band data, since these instruments are less sensitive
to small changes in the scattering characteristics.
Presenting analytical models for temporal decorrelation has proven to be unsuccess-
ful so far, since the range of possible temporal changes is too wide. For example,
anthropogenic changes caused by farming activities or construction works cannot be
modeled quantitatively due to its unpredictable and often discrete nature. Specific
weather conditions can have identical effects. On the other hand, for temporal decor-
relation caused by natural processes such as vegetation growth or seasonal effects
rules of thumb might be applied. For example, complete decorrelation due to snow
cover might change to good correlation as soon as the snow disappears. Experiments
with 1-day and 3-day period interferograms show that rates of decorrelation differ
considerably. Assuming that the temporal decorrelation is caused by the movement
of scatterers only, Zebker and Villasenor (1992) obtain an approximation based on
the rms of the displacements.
Quantitative analysis of temporal decorrelation, for example for classification pur-
poses, is not straightforward. Additional effects such as geometric decorrelation and
volume decorrelation need to be quantified separately in order to obtain meaningful
measures (Hoen and Zebker, 2000). Volume decorrelation, related to the penetration
of the radar waves, depends significantly on the radar wavelengths and the scattering
medium, see Hoen and Zebker (2000) for a review.

4.4.5 Interpolation errors

One of the first steps in SAR interferogram processing is the resampling of one
complex SAR image to map it onto a second image to within an accuracy of
about a tenth of a resolution element.
 4.4 Sources of decorrelation 105

Although the implementation may be different, resampling can be viewed as con-

sisting of two steps:

1. reconstruction of the continuous signal from its sampled version by convolution

with an interpolation kernel
2. sampling of the reconstructed signal at the new sampling grid.

This scheme holds even in many cases, where the convolution [step 1] is not obvious.
For example, nearest neighbor and Lagrange-type interpolation of equidistantly sam-
pled data can be considered as a convolution with particular kernels. The choice of
the interpolation kernel (especially its length) requires a tradeoff between interpola-
tion accuracy and computational efficiency. This shows that straightforward
system theoretical considerations give objective criteria for choosing or designing in-
terpolation kernels for interferometric processing.

Theory of interpolation errors The following analysis starts from the classical
Fourier domain description of interpolation errors in stationary signals. Often these
errors are quantified in terms of an We will instead employ coherence
theory for SAR interferograms (Just and Bamler, 1994; Cattabeni et al., 1994) to
predict the effect of interpolation on interferogram phase quality. Figure 4.10A shows
(for the one-dimensional case) how the Fourier transform of a kernel acts as
a transfer function on the periodically repeated signal power spectral density
The two classes of errors to be considered are the distortion of the useful spectral
band and the insufficient suppression of its replicas
where is the sampling frequency. Hence, the interpolated signal will not be strictly
low-pass limited and the subsequent new sampling creates aliasing terms. If in the
resampling process all interpixel positions are equally probable, the aliasing terms
are superposed incoherently and can be treated as noise with a signal-to-noise ratio
of:

In the following we will quantify interpolation errors in terms of interferogram decor-

relation and associated phase noise. We assume that the original data have been
sampled at least at the Nyquist rate and that the sampling distance after resampling
is similar to the original one.
The system model of fig. 4.10B is sufficient for our analysis: consider a perfect
and noise-free interferometric data pair of coherence before interpolation.
Both and have passed the SAR imaging and processing system described by a
transfer function Signal additionally suffers from the interpolation transfer
function I( f ) and alias noise n. It can be derived from Just and Bamler (1994)
This section has been published by Hanssen and Bamler (1999), and is here extended with the
analytic Fourier transform of the 6-point cubic convolution kernel.
106 Chapter 4: Stochastic model for radar interferometry

and Cattabeni et al. (1994) that, for circular Gaussian signals (i.e., for distributed
targets), the coherence of such a system is given by

These equations are readily extended to two dimensions, azimuth and range. If both
and are separable, we find:

The phase noise resulting from is known to be (in the N-look case):

where

Examples for interpolators The interpolators and their spectra evaluated here
are (assuming unity sample grid distance) as follows (Keys, 1981; Park and Schowengerdt,
1983).

Nearest

Note that the rect function needs to be 0.5 at the discontinuities to meet the Dirichlet derivative
conditions (Bracewell, 1986). In practice, however, it is often defined asymetrically.
 4.4 Sources of decorrelation 107

Piecewise linear interpolation:

Four-point cubic convolution

Six-point cubic convolution

Truncated sinc

where

Table 4.3 lists the theoretically derived coherence and one-look phase noise intro-
duced by the first three and the last of these interpolators for one and two dimensions.
ERS range signal parameters have been used for both dimensions with uniformly
weighted spectrum and oversampling ratio of
108 Chapter 4: Stochastic model for radar interferometry

(in real systems, azimuth oversampling is slightly higher than in range), i.e., H(f) =
rect(f / B). Often, SAR data are oversampled by a higher factor before an interfero-
gram is computed, be it either to avoid undersampling of the interferogram or as a
consequence of baseline dependent spectral shift filtering. In these cases the require-
ment on the interpolator is relaxed. Figure 4.11 shows how decorrelation reduces
with oversampling.

Experimental results Using uniformly-distributed random generators a 1-D

white circular Gaussian signal is computed with amplitude and
phase arg Low pass filtering yields a correlated random signal. An
oversampling ratio of 12.23 is used to create the reference signal u, whereas the test
signal is a subsampled version thereof. Using a subsampling ratio of 1/10 reduces
the oversampling ratio of the test signal to 1.223 to resemble ERS conditions. The
test signal is then interpolated using the kernels under investigation, yielding
an estimate of the reference signal. The interpolation kernels nearest neighbor,
piecewise linear, four-point and six-point cubic convolution and six-point, eight-point,
and 16-point truncated sine are created using eqs. (4.4.22a)–(4.4.26a). For every
kernel the interferometric phase error the phase error histogram, the
total coherence and the standard deviation of the interferometric phase error
are evaluated. Single experiment results of the interferometric phase error are shown
for the first four evaluated kernels in fig. 4.12. It can be seen that the variation of the
interpolated signal decreases considerably as the kernel contains more sample points.
Nevertheless, spurious spikes up to still cause residues in the interferogram. The
histogram is depicted in fig. 4.13A. The total coherence and the standard deviation
of the interferometric phase is studied using averaged values from 500 simulation
loops. The results are given in Table 4.3 to allow comparison with the theoretical
findings. Coherence has been estimated as the sample correlation coefficient of the
reference signal u and the interpolated signal Figure 4.13B shows the mean
 4.4 Sources of decorrelation 109

standard deviation of the phase as a function of the coherence for the four shortest
kernels.

Discussion of results The spurious spikes in the interferogram, as shown in

fig. 4.12, appear at those positions in the signal where the amplitude is extremely low.
The signal-to-noise ratio at these interpolation points is therefore dominated by the
interpolation noise. This makes a sudden phase jump at low-amplitude areas likely.
Due to the small amplitude, multilooking suppresses these spikes and considerably
diminishes the phase noise.
The cubic spline interpolation kernels used here can be referred to as parametric
cubic convolution (PCC) (Park and Schowengerdt, 1983). The parameters and
for the four- and six-point kernels chosen here have proved to be close to optimal for
this particular configuration. Optimization for specific purposes can be performed
by evaluating eq. (4.4.17).
A comparison of cubic splines and truncated sines underlines the necessity of careful
interpolator design: note that the cubic splines used here are special cases of weighted
truncated sines. The six-point cubic convolution kernel showed better quality than
an eight-point unweighted truncated sinc.
Interpolation errors are due to the aliasing of repeated signal spectra and the cutoff
of the signal spectra’s corners. Hence, the choice of an optimal interpolator will
always depend on the correlation properties of the signal. However, a subjective
recommendation can be given for ERS conditions, where temporal decorrelation
dominates the interferogram quality anyway. In these cases, a four-point cubic
convolution with = – 1 proved to be sufficient. For high resolution applications
of high coherence single-pass interferometers, in which multilooking is not desirable,
longer interpolation kernels, like the optimized six-point cubic convolution presented
here, are recommended.
110 Chapter 4: Stochastic model for radar interferometry

4.4.6 Coregistration errors

Coherent interferometric SAR products rely on an optimal coregistration of the two

complex images. Errors in the alignment introduce loss of coherence, and therefore
phase noise. It is obvious that a shift of a full resolution cell results in complete
decorrelation for distributed scattering mechanisms, since there is no physical cor-
respondence between the scatterers in both images left. Therefore, subpixel coregis-
tration accuracy is necessary to obtain coherent interferometric products. A relative
shift as a fraction of the range resolution results in a coherence of, see Just
and Bamler (1994):

This equation holds for a range as well as an azimuth misregistration. Application

of this equation results in fig. 4.14, where the effect of a misregistration is visu-
alized for coherence and phase standard deviation. Note that in this diagram it
 4.5 Integer phase ambiguities 111

is assumed that both signals are identical and only misregistration in-
fluences the coherence. The top axis shows the pixel cell size (posting), which is
approximately a factor 1.22 times the resolution cell size, both in range and azimuth
direction. The figure shows that the coherence does not improve significantly when
the coregistration accuracy is better than 0.1 resolution cell (l /8th of a pixel). For
this coregistration accuracy we find a coherence and a phase stan-
dard deviation of Since an identical result can be obtained in the azimuth
direction, we find the total coherence as a result of the coregistration to be
the product of both directions:

4.5 Integer phase ambiguities

An analysis of the quality, efficiency, and statistics of phase unwrapping algorithms

is outside the scope of this study. Such analyses or reviews can be found, e.g., in
Fornaro et al. (1996b); Ghiglia and Pritt (1998); Zebker and Lu (1998), or Bamler
and Hartl (1998). Nevertheless, some general remarks are in order to stress the
importance of phase unwrapping and its consequence for the quality of the final
results.
The phase statistics discussed in the previous sections can be regarded as single-
point statistics. They provide estimates for the phase variance—either through the
coherence or through the SNR—but do not influence the expectation value of the
phase data. In contrast, if only a single SAR interferogram is available, phase un-
wrapping is only possible by considering the phase differences between neighboring
pixels, making it essentially a multiple-point statistic. Errors induced by the integra-
tion of the phase gradients in one part of the interferogram can propagate into other
parts easily. Therefore, the quality of the solution at a certain location cannot be es-
112 Chapter 4: Stochastic model for radar interferometry

timated only by analyzing its direct surroundings. Figure 4.15 shows a section of the
probability density function of the absolute phase, which is known up to a ambi-
guity. Due to this ambiguity, the absolute position of the PDF is unknown, showing
the integer nature of the unwrapped phase solutions. If the expectation value of the
wrapped phase is denoted as the expectation value of the unwrapped phase
should be Figure 4.16 shows an example of phase
unwrapping errors, caused by a least-squares algorithm which doesn’t account for
the integer nature of the phase data. The result of phase unwrapping (fig. 4.16A)
is shown in fig. 4.16B. Although the unwrapping seems successful—the unwrapped
phase looks continuous—“spikes” in the result (e.g., at the white arrows) indicate
residues that have not been unloaded. If the algorithm would obey the
condition, such spikes would not show up. Moreover, noisy areas such as the lake
in the center of the image have been simply interpolated, hereby loosing all possible
phase information. Another indication of failing the is shown in
fig. 4.16C, where the unwrapped phases are “re-wrapped” to the interval,
and subtracted from the original data (fig. 4.16A). The would imply
that this difference needs to be 0 rad (gray) for the full image. It is clearly visible
that some areas have been biased after unwrapping. Areas with decreased coherence,
such as the hills in the south-western part of the image, have been interpolated.
Fortunately, most of the currently used algorithms obey the integerness-condition.
This implies that the errors introduced by phase unwrapping have a discrete
distribution. It is obvious that this characteristic of the data constrains a proper
quality description of the final interferometric products (topography/deformation),
in the sense that the PDF of these products is ambiguous as well.
 4.6 Influence and modeling of orbital errors 113

4.6 Influence and modeling of orbital errors

Spaceborne synthetic aperture (SAR) interferometry requires orbital satellite ephe-

merides or state vectors in order to determine the interferometric baseline vector
and to refer the interferometric products to a reference datum. As such, orbit errors
propagate directly into errors in topographic height or deformation maps. For global
applications such as potential field modeling or oceanography, orbit errors are conve-
niently parameterized as “once per revolution” errors, and much effort is expended
to improve gravity field models, dynamic models, and the quality of geodetic track-
ing measurements. The way orbit errors propagate into SAR interferograms differs
from this conventional approach, in the sense that the absolute orbit accuracy is less
important than the relative accuracy between the state vectors during the few sec-
onds of image acquisition. The required absolute accuracy for orbit determination
would need to be on the order of 1 mm to fully correct the residual interferometric
fringes, which is far below the current value of around 5–10 cm.
In the following it is described how orbit errors affect the interferometric baseline
and how a baseline error influences the so-called “reference phase,” that is, the
expected interferometric phase for a reference body such as an ellipsoid. After
a brief review of precise orbit determination in section 4.6.1, the interferometric
geometry is discussed in section 4.6.2. In the range direction the reference phase
can be expressed as a function of the baseline geometry and the range pixel number,
which is described in section 4.6.3. Section 4.6.4 covers how the predicted errors in
the state vectors propagate into baseline errors and residual reference phase signal.
Different methods to model and reduce the residual reference phase are discussed in
section 4.6.5. This section also evaluates the remaining errors after applying different
correction methods.
114 Chapter 4: Stochastic model for radar interferometry

4.6.1 Precise orbit determination

For ERS-1 and ERS-2, precise orbit determination (POD) is based on tracking data,
a gravity model, and several dynamical models. Tracking data include satellite laser
ranging (SLR) for ERS-1 and a combination of SLR and data from the Precise Range
and Range-Rate Equipment (PRARE) for ERS-2. As the tracking data from SLR
can only be acquired in the vicinity of an SLR station, which are unevenly distributed
over earth, gravity field information (a dynamic model) is needed to determine or
interpolate the position of the satellite at other positions. Gravity models, such
as GEM-T2, PGM035, JGM 3, and DGM-E04 are used for this purpose. Over the
oceans, altimeter crossover height differences can be used as additional tracking data.
Dynamical models are used to include nonconservative forces such as atmospheric
drag and solar radiation. Precise orbit determination incorporates a data adjustment
using observations and models, and produces a set of state vectors consisting of time,
velocity vectors, and position vectors (Scharroo and Visser, 1998).
Geographically correlated orbit errors, such as errors in the gravity field model, are
repetitive along the same ground track (Scharroo and Visser, 1998). For repeat-pass
SAR interferometry, where collinear tracks are a requisite, such errors are identical
and cancel. The remaining error sources include nonconservative forces such as
time variable atmospheric drag, time variable solar radiation pressure, and other
unmodeled effects.

4.6.2 Interferometric baseline geometry

The reference phase variation is due to the geometric configuration of the two satel-
lites and the variation of this configuration in the azimuth direction of the SAR
image. The geometric configuration is sufficiently modeled by the interferometric
 4.6 Influence and modeling of orbital errors 115

baseline vector, which can be described conveniently using three different represen-
tations, see fig. 4.17A the parallel/perpendicular representation, fig. 4.17B the hor-
izontal/vertical representation, and fig. 4.17C the representation in baseline length
and angle, written in vector notation as

see fig. 4.17. The parallel/perpendicular representation can be readily used to de-
scribe the sensitivity of the configuration for topographic heights and the amount of
geometric (baseline) decorrelation. The perpendicular baseline, can be regarded
as a scaling factor for the height ambiguity. A large usually implies geometric
(baseline) decorrelation, which decreases the SNR of the interferogram. The parallel
baseline, is in fact the basic path length difference

that is inferred directly from the absolute interferometric phase. and cor-
respond with the geometric range between a resolution cell on earth and antenna
positions 1 and 2, respectively. An important aspect of the parallel/perpendicular
representation is its dependence on the look angle (fig. 4.17A). This implies that
both baseline parameters depend on the relative position of a pixel in the inter-
ferogram, which is related to (i) orbit convergence/divergence, (ii) the geometric
difference between near-range and far-range, and (iii) the contribution of topogra-
phy.
In contrast to the parallel/perpendicular baseline, the horizontal/vertical represen-
tation is equal for all pixels in the same range line, as it is independent of the look
116 Chapter 4: Stochastic model for radar interferometry

angle (fig. 4.17B). The components are written as and respectively. The
advantage of this representation is that the horizontal baseline can be directly cou-
pled to across-track orbit errors, while the vertical baseline is directly related to the
radial component of the state vector. This representation will be used in the error
analysis in section 4.6.4.
The third baseline representation—in terms of baseline length B and orientation
can be useful because it can be easily related to the look which is necessary
to explain certain peculiar orbital phase effects, see the examples in section 4.6.3.
Sign of the baseline components
Regarding the sign of the baseline parameters we need to account for the interfero-
gram formation equations. Let and be the complex values of image 1 and 2.
In terms of amplitude and phase components we can write:

with Here we only regard the geometric part of the phase observation,
hence no scattering components. The interferogram is formed by complex multipli-
cation:

and therefore the relation between the interferometric phase and the
range difference to both satellites is

The range difference is equal to the parallel baseline see eq. (4.6.2),
so for as sketched in (fig. 4.17A), has a negative value.
The sign of is defined indirectly by the definition of orientation angle Here
we define α positive counterclockwise (consistent with the look from the
reference satellite (number 1), starting from the horizontal at the side of the look
direction (or from the horizontal opposite to the look direction for satellite 2, as in
fig. 4.17C). In this definition, is positive whenever satellite 2 is located to the
right of the slant-range line of satellite 1. For convenience: when then
will increase from near-range to far-range, or from the foot to the
top of a mountain, which results in a decreasing phase. In table 4.4, the conversion
formulas between the three representations are listed.
Knowledge of the baseline parameters is necessary to predict the interferometric
phase variation for an arbitrary reference surface. We can subtract this reference
phase from the observed data, and interpret the phase residuals as, e.g., topographic
height, deformation, or atmospheric signal. Frequently used reference surfaces are
locally best fitting ellipsoids, so that ellipsoidal heights can be obtained, but in fact
any suitable reference body is possible, even a geoid or a sphere.
 4.6 Influence and modeling of orbital errors 117

Ideally, in order to predict the reference phase, we would need to determine

for every azimuth and range resolution cell i and j. However, often it appears that
the variation of with time, denoted by can be approximated by a linear
drift for the acquisition period of a SAR image, typically 15.4 seconds for ERS-1/2.
In range direction, the phase of the reference surface can be predicted by directly.
Thus, we need only and to compute the reference phase in two dimensions.

4.6.3 Reference phase in range

To derive a convenient analytical expression for the reference interferometric phase,

a spherical approximation is used, see fig. 4.18, where the earth radius is adapted
to the latitude of the spacecraft. The slant range will be used as input parame-
ter, depending on the range pixel number. Using straightforward trigonometry, an
expression for location angle is derived, which is related to the interferometric
reference phase
The horizontal and vertical baseline components are readily derived from the state
vectors and using

The slant-range value is derived from the roundtrip travel time to the first range
pixel, the sampling frequency and the velocity of electromag-
netic waves c. The range increments dR can be found using

and the range to the first range pixel is

118 Chapter 4: Stochastic model for radar interferometry

Using (4.6.7) and (4.6.8) the range as function of pixel i can be expressed by

The relation between the measured range and the unknown location angle is

and, using the same configuration for the second satellite position, we find

using the far-field approximation—assuming parallel state vectors and In

range direction, and are constant. Therefore, as changes from
near-range to far-range, only will change. For a coregistered image, we want to
express the values of as a function of From the interferometric measurements
we find that, cf. eq. (4.6.5),

Combining eqs. (4.6.11) and (4.6.12), we find for the reference phase
 4.6 Influence and modeling of orbital errors 119

and from (4.6.10) follows

Note that in fig. 4.18, is largely exaggerated: in reality is only a few degrees.
Using

and

we find an expression of as a function of only:

where

Using (4.6.9), where equation (4.6.17) describes the behavior of the

reference phase for range pixel i. The reference phase depends on the baseline
vector here expressed using the components and In the next section, it
will be shown that this parameterization enables a direct relation to the radial and
across-track orbit errors.
Figure 4.19 shows three baseline configurations with their corresponding reference
phase in range direction. The figure shows how the phase variation is directly related
to the parallel baseline. Moving from near-range to far-range
24°) increases in fig. 4.19A and 4.19B. Since values are positive-
decreasing in fig. 4.19A and negative-decreasing in situation fig. 4.19B. Obviously,
reversing the two satellite positions will yield increasing values in both situations.
Although the reference phase variation seems nearly linear in fig. 4.19A and 4.19B,
a special situation occurs when the baseline orientation is in the range

as shown in fig. 4.19C. In this interval the reference phase will be both increasing
and decreasing. The right column of fig. 4.19 shows the absolute phase in radians
as a function of the range direction corresponding to the sketches in the left
column. The insets show the corresponding wrapped phase equivalents.
Note that the behavior of the reference phase (increasing or decreasing) is caused by the
behavior of the parallel baseline The latter, however, can also be directly derived from the sign
of the perpendicular baseline
120 Chapter 4: Stochastic model for radar interferometry
 4.6 Influence and modeling of orbital errors 121

4.6.4 Description of orbit errors

Errors in the orbital state vectors can be represented as errors in the along-track, the
across-track, and the radial directions, as sketched in fig. 4.20A. For SAR interferom-
etry, along-track errors are usually sufficiently corrected for during the coregistration
of the two images, see fig. 4.20B. In fact, it is sufficient to perform the alignment of
the images with an accuracy of 1/10th of the resolution size, see section 4.4.6, which
results in a derived along-track accuracy of approximately 40 cm for ERS. Note that
along-track positioning errors can also be regarded as timing errors.
As only radial and across-track errors will propagate as systematic phase errors in
the interferogram, the problem is effectively two-dimensional. From the SAR image
coordinates, range and azimuth pixel positions, these effects can be separated in a
nearly instantaneous component in range direction, and a time-dependent compo-
nent in azimuth direction. The effects will be discussed in both directions.
Errors in the two state vectors need to be propagated to the baseline vector rather
than analyzed separately, since an interferogram is an inherently relative measure.
A quasi-3D plot is shown in fig. 4.21 A, indicating the change in the baseline vector
and the influence of radial and across-track orbit errors. Using the DEOS precise
orbits, radial and across-track rms errors are on the order of 5 and 8 cm respectively
(Visser et al., 1997; Scharroo and Visser, 1998). Although geographically correlated
errors cannot be resolved using this approach—which makes the estimate often too
optimistic in an absolute sense—this is of no consequence for the interferometric
baseline.
The relationship between baseline vector errors and radial/across-track errors is
122 Chapter 4: Stochastic model for radar interferometry

given by:

assuming uncorrelated errors between orbit 1 and 2. This implies that the orbit er-
rors can be regarded as a single “noise vector” superposed on the baseline estimates
as sketched in fig. 4.21B.
Representing orbit errors as a noise vector on top of the baseline vector leads
to a convenient way of interpreting residual orbit fringes. The noise vector can
be regarded as a second baseline vector. As we subtract the expected (a priori)
reference phase, a residual reference interferogram will remain, with the unknown
a posteriori “baseline” vector Unfortunately, the only available information we
have on is that its length is less than two times the largest value of and
with a probability of 95%. However, as no information is available on the orientation
of denoted by in fig. 4.21B, the residual reference interferogram can have any
kind of behavior. Substituting B, and by n, and respectively in fig. 4.19
illustrates this relation. Obviously, since n B, the amount of fringes decreases
considerably. Figure 4.22 illustrates the behavior of the reference phase in range
direction for a radial rms of 5 cm and an across-track rms of 10 cm, for a range of
angles Experiments with similar radial and across-track rms values show that
the largest effects occur when whereas when the error can
be almost neglected. However, since there is no way of knowing it seems wise to
account for the largest errors. In fig. 4.23 the maximum amount of residual fringes
in an interferogram is shown, as a function of the a priori radial and across-track
rms values of a single state vector. With a 95% likelihood, the graph shows the
maximum amount of residual fringes in a 100 x 100 km interferogram.
 4.6 Influence and modeling of orbital errors 123

From this relation, the approximate orbit accuracy can be derived from the number
of residual fringes in the interferogram. One fringe or less corresponds to rms values
of 5 and 10 cm in radial and across-track direction, and keeping the ratio between the
two error sources constant, the errors scale nearly linear with the number of fringes.
For example, when observing 16 residual fringes in the interferogram, which cannot
be contributed to, e.g., topography, expected maximal rms values are approximately
80 and 160 cm in radial and across-track direction.

4.6.5 Correcting the residual reference phase

Knowing how residual fringes originate from orbit errors, it needs to be studied how
they can be eliminated from the interferogram. Strictly, the only way to do this is to
try to estimate the total orbit error vector The parallel component is needed
to correct for the residual reference phase, whereas the perpendicular component
124 Chapter 4: Stochastic model for radar interferometry

needs to be added to the perpendicular baseline to obtain the correct height

conversion factor.
If the interferogram is used for topography height estimation, it seems suitable to use
tie-points to constrain the reference phase at selected locations in the interferogram.
A similar approach is used by Massonnet and Feigl (1998), see fig. 4.24A. In this
approach, one satellite position is fixed, while the second one, indicated in the figure,
is changed in two steps. First, since the range distance to the first pixel in
range can be kept constant, the position of the satellite needs to be somewhere on
curve 1. By counting the number of fringes between near-range and far-range, range
vector needs to be changed in length, denoted as step 1. Keeping this value for
the position of the satellite needs to be somewhere on curve 2. Therefore, in
step 2, the position is moved to the cross-section of curve 1 and 2. Using the same
approach in azimuth direction, residual azimuth fringes can be eliminated.
Although it seems that this approach solves for the error vector this is in fact
not totally true. In fact, the method corrects only for the change in the parallel
component of the error vector, see equation (4.6.20), which causes the residual
fringes in the interferogram. The method does not correct for the perpendicular
component of the error vector, and still leaves an error in the height conversion
factor. In fig. 4.24B it is shown how this method fails finding the correct position of
the satellite. The a priori position of the satellite is indicated as position 1. The true
position of the satellite is indicated as position 2. The error vector points from
position 1 to 2. As B is often much larger than n, the major part of the reference
phase is decreasing with range, indicated by the boldly drawn parallel component of
The reference phase from this baseline can be eliminated totally. The influence
of the error vector is not known, though. From the figure it is clear how is
also decreasing with range. The length change of from near-range to far-range is
directly related to the number of residual fringes in the interferogram. This length
 4.6 Influence and modeling of orbital errors 125

change will be denoted as

Since is unknown, we can only observe the residual fringes in the interfer-
ogram, which are equivalent to In fact, we observe from the interferogram
that is too short by Correcting for this error is indicated as step 1
in fig. 4.24B. Keeping constant, the position that Massonnet and Feigl (1998)
would find will be the intersection of curve 1 and curve 2, indicated as position 3.
It is clear from the figure that this method might eliminate the residual reference
phase, but it decreases the perpendicular baseline whereas the correct solution
increases With an angle of approximately 6° between the near-range and far-
range look angle, a correction for two residual fringes (approximately 6 cm) would
lead to a decrease in of 57 cm. Such an error would result in an error of 53 cm
per fringe for a baseline of 100 meters, and 9 to 56 meters per fringe for a baseline
of 25 to 10 meters, respectively.
A different problem connected with using tie-points is caused by atmospheric phase
variation. To translate the height difference between two tiepoints to a phase dif-
ference, it has to be assumed that there are no additional phase gradients in the
interferogram, for example caused by atmospheric heterogeneities. Previous stud-
ies have shown that the phase difference between a set of tiepoints within a SAR
image can easily exceed several fringes. Thus, without any prior information on
atmospheric behavior during the image acquisitions it seems cumbersome to correct
for residual reference phase. It could be tried to diminish this problem by using a
large number of equally spread tiepoints over the whole image, assuming that at-
126 Chapter 4: Stochastic model for radar interferometry

mospheric phase signal will average out in the adjustment procedure. However, for
interferograms which have deformation as well as topographic signal, the tiepoint
approach cannot be used.
In practical applications, a number of phase interpretation problems accumulate.
Phase gradients can be due to atmospheric signal, errors in the topographic elevation
model, or long wavelength deformation signal. In some cases, the specific behavior of
the phase gradients enables identification of the driving mechanisms. Unfortunately,
for largely decorrelated interferograms as used for this study, phase gradients cannot
be followed over an extended spatial range. Therefore, they can be due to any of
the mentioned sources. Moreover, phase unwrapping between the coherent patches,
here referred to as “patch unwrapping,” can cause severe misinterpretation of phase
gradients, see section 5.4.
For this reason it is valuable to have a first order approximation of the type, behavior,
and magnitude of residual reference phase errors, only based on the a priori radial
and across-track orbit accuracies. In range direction, the residual reference phase is
approximated by a linear and quadratic polynomial. The linear model corresponds
with fitting a linear plane through the interferogram. Using the configuration shown
in fig. 4.21B, and the a priori radial and across-track orbit rms errors, a set of
noise vectors is simulated, with lengths corresponding to the 95% level under 192
equally spaced orientation angles For each of these combinations the values of the
polynomial coefficients for the 1st and 2nd degree approximations are determined, as
well as the error between the true signal and the approximation. The error is defined
as the maximum residual phase value minus the minimum residual phase value. The
results are shown in fig. 4.25. It can be observed that using a linear correction for
the residual reference phase, a maximal phase error of approximately 0.35 rad (6%
of a cycle) remains in the interferogram after correction. These maximum errors
occur when the noise vector is nearly horizontal. Scaling the radial and across-
track rms errors with a factor f will linearly scale the maximum error. Since the
linear trend is removed, the shape of the remaining error in the interferogram will
be approximately hyperbolic. From this distribution, without knowledge of the
probability of a residual error larger than a certain threshold can be obtained.
Using the quadratic polynomial approximation, the ratio between the coefficient of
the linear term and the coefficient of the quadratic term is approximately
indicating that the signal is only weakly curved. This value is determined by the
sensor height and viewing direction. The error approximated with the quadratic
polynomial approach is shown in the lower right panel of fig. 4.25. This error is a
scaled-down version of the linear approximation error with a factor 0.1. Maximal
errors reach 0.035 rad (0.6% of a cycle). The shape of this small remaining error is
a sine-like curve over the whole range line.
The severity of remaining errors depends on the application. For example, for topo-
graphic mapping using a small perpendicular baseline the residual error after
correction with a linear trend can be significantly (6% of the height ambiguity). In
those cases, it can be recommended to remove a quadratic polynomial surface from
the interferogram. It has to be noted that the ratio between the magnitude of the
first and the second degree coefficient in this case has to be restricted to approx-
 4.6 Influence and modeling of orbital errors 127

imately when approximating these coefficients from the interferogram. The

shape of the error signal in fig. 4.25 is dependent on the ratio between the radial and
across-track error. For equally sized errors, the curve can be described by
A relatively larger across-track error will make the curves more pointed.

4.6.6 Note on the reference system

A spherical approximation has been used for convenience in obtaining analytical

equations. The equations governing the behavior of the reference phase can, however,
be easily extended to other reference systems, such as ellipsoidal systems or the geoid.
In order to do this, the term expressing the earth radius needs to be replaced by,
e.g., where H expresses the height of the chosen reference surface over
the spherical earth. This height might be expressed in a spherical harmonics series,
as a function of latitude and longitude. For example, using the geoid as reference
surface, the height N between ellipsoid and geoid needs to be known, as well as
the height between that ellipsoidal position and the sphere. Both height levels
128 Chapter 4: Stochastic model for radar interferometry

are not parallel, as the mass distribution of the gravity anomalies cause deviations,
see fig. 2.9 at page 33. Nonetheless, since an interferogram is a relative measure,
these effects will only be of influence when the error due to this approximation is
significantly different between, e.g., the two opposite sides of the interferogram. For
example, the total bias between, e.g., ellipsoidal height with respect to the sphere
is not important: only gradients can be measured.

4.6.7 Conclusions
Errors in the orbital state vectors result in a residual reference phase in the radar
interferogram, which can be erroneously interpreted as topography, deformation,
or atmospheric signal. These errors are not related to the baseline length B and
orientation angle In fact, it is not possible to predict whether the residual reference
phase is increasing, decreasing, or both in range direction, as this is dictated by the
unknown orientation of the error vector The total residual reference phase is
restricted to the range where is determined by
the magnitude of the radial and across-track orbit errors.
Correction of the residual reference phase can be based on tiepoints, but this is
strongly dependent on the amount and extent of atmospheric or deformation signal in
the interferogram. Especially for interferograms with extended decorrelated regions,
this technique cannot be applied. A method proposed by Massonnet and Feigl (1998)
can be used for eliminating the residual reference phase, but it may introduce new
errors in the perpendicular baseline
Approximations of the residual reference phase, based on a priori estimations of the
radial and across-track orbit errors, are valuable to suppress the residual reference
phase, and to assess the magnitude and behavior of the maximum error after correc-
tion. For rms values of 5 and 10 cm in radial and across-track direction respectively,
it is found that a linear approximation will suppress the residual reference phase to
a maximum error of 0.35 rad, in a hyperbolic shaped plane. A quadratic polynomial
model further decreases the maximum error to 0.035 rad, in a sine-like plane. In or-
der to apply a quadratic polynomial model, as a rule of thumb, the ratio between the
magnitude of the second degree term and the first degree term needs to be restricted
to

4.6.8 Propagation of orbit errors to height errors

Assume that the orbit errors, expressed as an rms in radial and across-track di-
rection, result in an error baseline with length and orientation see
fig. 4.21B. Since the topographic height is related to the interferometric phase with,
see eq. (3.2.1),

where
 4.6 Influence and modeling of orbital errors 129

see fig. 4.21C. Here, is the perpendicular baseline derived from the observed state
vectors and is the residual (error) perpendicular baseline caused by the error in
the state vectors. Hence the topographic error is

where is the initial height derived using the available orbit information. Hence,
the baseline error results in an erroneous scaling of the topographic height with
factor For perpendicular baselines larger than 20 m, the scaling factor
is typically less than 1%, using ERS precise orbits. To derive the absolute error in
an interferogram with an arbitrary state vector accuracy and perpendicular baseline
we need to evaluate the propagation of the radial and across-track orbit PDF to the
PDF of the perpendicular error baseline
Assuming a Gaussian PDF for the radial as well as the across-track orbit error, see
eq. (4.6.19a), we find the joint probability density function for the baseline error to
be

where and cf. eq. (4.6.19a). In fig. 4.26A the marginal

PDF’s for the horizontal and vertical baseline errors are shown in combination with
the joint PDF, based on a 5 cm radial rms and an 8 cm across-track rms. A
simple coordinate transformation from the horizontal/vertical baseline error to the
perpendicular/parallel baseline error can be written as:
130 Chapter 4: Stochastic model for radar interferometry

and the marginal PDF for is consequently the integral over all parallel baselines

As a result, the probability of an error perpendicular baseline to be less or larger

than a certain threshold value can now be numerically derived, based on the rms
values of the state vectors. In fig. 4.27A the probability of an error baseline to be less
than a threshold in meters, is given. Figure 4.27B shows for four perpendicular
baseline lengths how this influences the scaling factor For example, for
there’s a probability of 70% that the erroneous scaling is less than
1%. Planning suitable SAR acquisitions for topographic mapping is possible using
eq. (4.6.26) or from fig. 4.27A. Suppose we want to map a mountain with an elevation
of 1 km with an accuracy better than 1 m. With a 90% likelihood this yields a
maximum perpendicular baseline error of 0.17 m, hence With eq. (4.6.23)
we find for dH = 1 m and that we need a perpendicular baseline

4.7 Atmospheric signal: turbulence

First indications that atmospheric delay variations might cause problems for the
interpretation of interferograms have been reported by Goldstein (1995); Massonnet
and Feigl (1995a); Tarayre and Massonnet (1996). There is currently no alternative
instrumentation to measure this delay with sufficient accuracy, spatial resolution,
and temporal sampling to eliminate its influence entirely from the interferograms.
 4.7 Atmospheric signal: turbulence 131

Neither is it possible to measure the refractivity distribution causing the delay,

cf. eq. (3.4.2), in three dimensions simultaneously with the SAR acquisitions. In
appendix A we show that attempts to eliminate atmospheric influence from interfer-
ograms using GPS (time-series or networks) can give a reasonable approximation of
the long wavelength features, but cannot eliminate signal with wavelengths shorter
than, say, 10 km. Moreover, it introduces interpolation errors in the modified inter-
ferogram.
Since it is not possible to correct for atmospheric errors in a deterministic way,
it is important to develop a mathematical model that describes the behavior of
atmospheric delay in interferograms stochastically. A stochastic model is necessary
for quality control, for example to prevent erroneous interpretation, and is imperative
for data adjustment using series of interferograms.
We distinguish two types of atmospheric signal, based on their physical origin.

Turbulent mixing, discussed in this section, results from turbulent processes in

the atmosphere. It causes spatial (3D) heterogeneity in the refractivity during
both SAR acquisitions, and affects flat terrain as well as mountainous terrain.
Vertical stratification, discussed in section 4.8, is the result of different vertical
refractivity profiles during the two SAR acquisitions, assuming there are no
heterogeneities within the horizontal layers. This affects mountainous terrain
only, and is correlated with topography (Massonnet and Feigl, 1998).

As vertical stratification is a function of the topographic height differences, which are

assumed to be unknown a priori, this subject is treated separately, see section 4.8.
In this section, we will focus on the effects of turbulent mixing.

4.7.1 Introduction

Three important observations characterize atmospheric signal in interferograms.

First, the relative character of an interferogram does not enable absolute signal
delay measurements. Hence, we can demean the interferogram, which implies that
the expectation value of the atmospheric signal for an arbitrary pixel will be zero. As
a consequence, there is no dependency on vertical variation in refractivity—different
vertical refractivity distributions can result in the same total delay. Second, or-
bit errors can easily cause a nearly linear trend over the whole interferogram, see
section 4.4.6. Such trends are usually hard to distinguish from atmospheric signal
delay trends, and are usually eliminated using some kind of residual flattening, for
example using tie-points. This strategy is equivalent to a high-pass filter for the
atmospheric signal. Finally, we know that the amount of atmospheric signal within
an interferogram is highly variable in time: some interferograms seem to have no
atmospheric influence at all, whereas others are contaminated significantly. This is
also dependent on the climatological characteristics of the location. Nevertheless, for
acquisition time intervals of 1 day or more the two states of the atmosphere appear
to be practically uncorrelated.
Several studies have shown that the predominant part of the atmospheric signal in
132 Chapter 4: Stochastic model for radar interferometry

interferograms is caused by the water vapor distribution in the lower troposphere

(Goldstein, 1995; Tarayre and Massonnet, 1996; Hanssen and Feijt, 1996; Tarayre,
1996; Zebker et al., 1997; Hanssen, 1998; Hanssen et al., 1999, 2001, 2000b). This is
a consequence of the fact that for microwave frequencies, the permanent molecular
dipole moment of dominates the variability of the refractive index. Moreover,
the water vapor distribution can be considered to be an approximate passive “tracer”
of the mechanical turbulence (Ishimaru, 1978). Therefore, the observed statistics
of the atmospheric signal will be correlated with those of wind vortices, see, e.g.,
Hanssen et al. (2000b).
The behavior of atmospheric signal in radar interferograms can be mathematically
described using several interrelated measures such as the power spectrum, the covari-
ance function, the structure function, and the fractal dimension. The power spec-
trum is useful to recognize scaling properties of the data or to distinguish different
scaling regimes. Such Fourier-domain representations enable elegant descriptions of,
e.g., filter operations and fast processing algorithms. A disadvantage is that Fourier
methods require data sampled on a regular grid (Chiles and Delfiner, 1999). The
spectrum of the grid is identical to the spectrum of the underlying continuous signal
only if frequencies higher than the Nyquist frequency do not occur in the signal, ei-
ther because they do not exist or because they are removed before sampling. Hence,
the bandwidth of the grid spectrum needs to be smaller than the Nyquist frequency.
For the SAR data this condition is usually satisfied.
The covariance function is theoretically equivalent to the power spectrum, although
its evaluation for irregularly spaced data can be easier and the signal does not have
to be bandlimited. It is used, e.g., to construct the variance-covariance matrix
which is a prerequisite for data adjustment and filtering. A disadvantage is that it
does not directly show the variance of the difference in signal between two points in
an image, its use is restricted to second-order stationary signals, and it requires an
a priori estimation of the mean value.
The structure function does not have this restriction and can be used for the broader
class of intrinsic functions, cf. fig. 4.1. This implies that it has no restrictions con-
cerning bandlimitedness, preliminary estimation of the mean, or stationarity. It
provides a useful quantitative expression for the variance of the difference in atmo-
spheric delay between two points with distance Since it cannot be directly used for
data adjustment procedures, it is useful for data quality description if there is only
one realization of the data. Finally, the fractal dimension provides an elegant mea-
sure for the roughness and scaling characteristics of the signal. In the following we
elaborate on the conversions between these measures, focusing on the atmospheric
signal in interferometry.

4.7.2 Theory

Turbulent mixing is a result of different tropospheric processes such as solar heating

of the earth’s surface (causing convection), differences in wind direction or velocity at
different layers, frictional drag, and large scale weather systems. Turbulent processes
cascade down from such large scales to smaller scales until the energy is dissipated.
 4.7 Atmospheric signal: turbulence 133

The range of spatial scales over which this cascading takes place is known as the
inertial subrange. Kolmogorov’s turbulence theory assumes that kinetic energy is
conserved within the inertial subrange. The turbulent wind vortices act as a carrier
for atmospheric constituents such as water vapor, hereby influencing the refractivity
distribution. Although the tropospheric refractivity for radio frequencies is mainly
dependent on temperature, pressure, and water vapor, it is dominantly water vapor
which causes the atmospheric signal within the dimensions of a SAR image, consid-
ering that only horizontal variability in refractivity is affecting the observed delay
in interferograms.
Since atmospheric delay due to turbulent mixing affects a wide range of scales,
governed by strongly nonlinear processes, there is a need to find the simplest and
most robust measures of the variability in the delay signal.
Kolmogorov turbulence theory
In the case of isotropic turbulence in three dimensions, Kolmogorov turbulence the-
ory predicts a specific structure function, of the spatial variation of the re-
fractivity N (Tatarski, 1961):

where and are the inner and outer scales of turbulence (Tatarski, 1961; Ishimaru,
1978; Treuhaft and Lanyi, 1987; Ruf and Beus, 1997). The first regime in eq. (4.7.1)
was developed by Kolmogorov (1941) and Obukov (1941), and is referred to as the
inertial subrange where the “two-thirds law” applies. The exponent 2/3 expresses
the rate at which the refractivity decorrelates with increasing distance. The structure
coefficient is a measure for the roughness of the spatial heterogeneities.
The inner and outer scales of turbulence have been studied in first instance for op-
tical signals. For those wavelengths, the inner scale is generally considered to be
in the order of a few millimeters whereas the outer scale is in the order of 5–10 m,
see Coulman and Vernin (1991) or Ruf and Beus (1997). For optical wavelengths,
temperature variations are the dominant factor of the refractivity variations (Lay,
1997). For radio wavelengths, the refractivity variations are dominated by water
vapor variability, see (Ishimaru, 1978). The outer scale for fully developed 3D Kol-
mogorov turbulence can be up to several kilometers, and is often indicated as the
effective height, scale height, or the tropospheric thickness (Dravskikh and Finkel-
stein, 1979; Armstrong and Sramek, 1982; Treuhaft and Lanyi, 1987; Coulman and
Vernin, 1991; Ruf and Beus, 1997).
The power spectrum where k is the wave number, within the inertial subrange
is (Tatarski, 1961)

for observations along a (one-dimensional) line through the three-dimensional vol-

ume, which is equivalent to time series at a point. For a two-dimensional cross-
section (slice) of the volume, using the radial wave number we find
134 Chapter 4: Stochastic model for radar interferometry

(Voss, 1988; Turcotte, 1997)

where the addition of minus one is required due to the radial coordinates. Another
addition of minus one is required when using spherical wave numbers for the volume.
In general, we can write

known as the Kolmogorov Power Spectrum. The wavenumbers and are

defined in an orthogonal Cartesian wavenumber frame. The relation between the
power spectrum and the structure function is discussed in following paragraphs.
Structure function of the signal delay
In fig. 4.28A the 3D refractivity distribution at the time of the first SAR
acquisition is sketched. Assuming that the radio wave source is high above
the turbulent layer, the 2D vertical cross-section in fig. 4.28B shows a flat, undis-
turbed wavefront, before entering the turbulent refractive layer. After propaga-
tion through the medium, arriving at at the earth’s surface, it is inhomogeneously
delayed. The atmospheric phase signal in an interferogram is the superposition of
this wavefront for the SAR acquisitions at and see section 3.4.
 4.7 Atmospheric signal: turbulence 135

The relation between the refractivity distribution and the one-way sig-
nal delay has been presented in eq. (3.4.2). Consequently, the theoretical struc-
ture function of the refractivity, eq. (4.7.1), can be propagated to the structure
function of the signal delay.
Assuming the Kolmogorov relation for isotropic 3D turbulence, within the inner and
outer scales, we used eq. (4.7.1), with constant over all distances Moreover,
in the calculation of the tropospheric delay covariance we will also assume that
is independent of the position i.e., second-order stationarity (Treuhaft
and Lanyi, 1987). The tropospheric delay for pixel k at location from a
satellite at incidence angle (zenith angle) see eq. (3.4.2), can be reformulated
as

where is the 2D location vector on the earth’s surface and h is the effective height,
see fig. 4.28A. is the vector between position and the satellite’s position.
For two positions with horizontal distance the expectation E for the quadratic
difference of the delay between the two pixels k and l, at positions and
is the structure function, c.f. eq. (4.1.8):

By substituting eq. (4.7.5) in (4.7.6), and exchanging the order of integration and
expectation (summation), we can express the structure function of the delay as
function of the refractivity (Treuhaft and Lanyi, 1987):

This derivation is discussed in appendix B.1. Assuming second-order stationary

signals we can write the structure function of N as

and the structure function of the delay as a function of the structure function
of the refractivity

Using eq. (4.7.1), the structure function of the delay behaves as (Tatarski,
1961; Thompson et al., 1986; Treuhaft and Lanyi, 1987; Coulman and Vernin, 1991):
136 Chapter 4: Stochastic model for radar interferometry

in comparison with the 2/3-exponent of the refractivity in eq. (4.7.1).

The power spectrum of the atmospheric delay signal, where k is the wavenum-
ber corresponding with the structure function in eq. (4.7.10), is (Tatarski, 1961)

The wavenumbers and are orthogonal wavenumbers in the plane.

Power spectrum of the signal delay
The general relation between the structure function and the power spectrum
of a random continuous one-dimensional signal with a power-law behavior is
discussed by, e.g., Monin et al. (1975) and Agnew (1992). If the power spectrum is
written as

where f is some (spatial or temporal) frequency, and are normalizing constants,

and is the spectral index (often The structure function of a signal
with the power spectral form of eq. (4.7.12) can be written as

The derivation of the relation between the power spectrum and the structure function
as well as the definition of the structure coefficient is given in appendix B.2.
The power-law model is a structure function for (Chiles
and Delfiner, 1999). Since this model doesn’t have a “sill” (a range at which the
variance of the difference between two points does not increase anymore), it would
need to be adjusted for atmospheric signal as observed by radar interferometry.
However, since the data ranges in an interferogram are spatially limited, and the
observations have a relative character, it is not necessary to expand the model for
ranges larger than, say 100–300 km.
Addition and scaling of power-law models
An interferogram contains the summation of the atmospheric signals at the two
acquisitions and Therefore, the power spectrum and structure function of
the signal in an interferogram is different from the signal observed by other space-
geodetic techniques, such as GPS or VLBI. Using the power spectrum of
(atmospheric) signal cf. eq. (4.7.12):
 4.7 Atmospheric signal: turbulence 137

where denotes the 1D Fourier transform, and the operation results

in the power spectrum. For the addition/subtraction of and the addition
theorem in Fourier transforms states (Bracewell, 1986)

Therefore, for a model atmosphere with Kolmogorov turbulence characteristics, the

observed power spectrum in an interferogram has the same scaling properties as the
atmospheric influence during only one SAR acquisition, indicated by the exponent
although the intensity of the variations is in fact the summation of the two
contributing power spectra.
Equivalently, using eqs. (4.7.13) and (4.7.16), the structure function of the summa-
tion of the two atmospheric signals is

Under the assumption that the atmosphere is statistically similar during the two
acquisitions we may write:

where the subscript “sar” indicates a single SAR acquisition, and the subscript “ifg”
indicates the interferogram.
The similarity theorem (Bracewell, 1986) is used when scaling a signal, e.g., when
transforming a GPS delay time series to spatial variation using the wind speed

This means that the same P(f) values will end up at a spatial wavenumber
scaled with factor Using double-logarithmic axes, this corresponds with a
simple shift of the curve, with identical exponent
Power spectrum and covariance function
According to the Wiener-Khinchine theorem, the power spectrum and the
covariance function of a second-order stationary process are each others cosine
Fourier transforms (Chiles and Delfiner, 1999):

with
138 Chapter 4: Stochastic model for radar interferometry

For a discrete finite signal with n = ( 1 , 2 , . . . , N) , the covariance

is derived for l = (1, 2 , . . . , N/2) using

The power spectrum is derived using the Discrete Fourier Transform (DFT):

and

where we only used the positive wavenumbers. If a function is the Fourier trans-
form of a nonnegative measure it is a positive (semi-)definite covariance function, as
proved by Bochner (1959).
Larger scales
Equation (4.7.1) has been defined for homogeneous isotropic 3D turbulence. For
horizontal scales commonly associated with, e.g., GPS and radar interferometry,
the spatial distances can be larger than the effective height of the wet troposphere
(1–3 km), hence, larger than the inertial subrange. This reduces the tropospheric
contribution to a relatively thin layer with effectively 2D turbulence. Stotskii (1973)
proposed to write the structure function of the signal delay for scales larger than
the effective tropospheric height h as

using eq. (4.7.1). For the power-law behavior of the power spectrum we find with
eq. (4.7.12) and (4.7.13):

where and are orthogonal wavenumbers in the plane.

Early experiments investigating atmospheric signal delay were performed using, e.g.,
VLBI antenna’s. Since these have a fixed position, hence fixed spacing, only few
observational data were available to verify Stotskii’s proposition. The InSAR data
analysis discussed hereafter provides independent support of this proposition.
The transition between the two regimes—3D-turbulence, eq. (4.7.11), to 2D turbu-
lence, eq. (4.7.27)—can be described in several ways. For example, Treuhaft and
 4.7 Atmospheric signal: turbulence 139

Lanyi (1987) used a polynomial to describe the transition fluently. Comparison with
the InSAR data suggests that such a polynomial form might be restricted to a spe-
cific location on earth. In this study, we define a distinct change between the two
regimes at scale height h. Since the spectral form is a stochastic representation of the
behavior of the signal, such an abrupt change does not introduce artifacts whereas
it is a convenient simplification.

4.7.3 Observations

To investigate the stochastic behavior of tropospheric signal in repeat-pass radar

interferograms a set of 26 interferograms over the Netherlands—corresponding with
52 weather situations—has been analyzed (Hanssen, 1998). The locations of the
interferograms are shown in fig. 4.29. Using a reference elevation model, differential
interferograms have been obtained, even though topographic variation is limited
to a 5–10 m range for the northern and southern interferograms, and maximally
140 Chapter 4: Stochastic model for radar interferometry

for the central part. ERS “tandem” imagery is used to minimize the effects
of temporal decorrelation and to exclude possible wide-scale surface deformation.
The interferogram data span the whole year, providing a reasonable sampling of the
seasonal variation. However, the imaging times are the same for all acquisitions,
prohibiting the study of variations between daytime and nighttime acquisitions.
Since orbit errors can introduce phase trends in the interferogram, the data are
detrended in azimuth and range using a linear polynomial model. Therefore, the
data can be considered second-order stationary. For these data the mean values
are zero by definition, but higher order moments vary considerably depending on
the weather situation, as shown in fig. 4.30. The observed rms values are shown
for an area of RMS values of 1.5 rad (6.7 mm slant delay) occurred
in most of the situations, almost 40%. Approximately 15% of the situations had a
low rms value of 0.5 rad (2.2 mm), whereas 10% of the situations had a significant
 4.7 Atmospheric signal: turbulence 141

atmospheric variation of 3.5 rad (15.6 mm) rms. Although these values give an
indication of the range of atmospheric contributions, they should always be related
to their spatial support. Therefore, other measures as discussed below provide a
better parameterization. Figure 4.31 shows an overview of the turbulent signal in
eight differential tandem interferograms, acquired over Groningen in the northern
part of the Netherlands during 1995/96. The interferometric phase is unwrapped
and converted to zenith delay signal in mm.
The state of the atmosphere varied from severe thunderstorms during the both ac-
quisitions contributing to one interferogram (fig. 4.31A) to extremely calm weather
without significant convection, e.g. fig. 4.31D. Hanssen (1998) provides a more elab-
orate meteorological description of these data.
Power spectrum and fractal dimension
The 50 x 50 km windows in the data of fig. 4.31 are used to calculate a set of
power spectra. First the mean value of the window is subtracted, and both in
azimuth and in range direction a linear trend is removed from the data. For all
eight interferograms of fig. 4.31, the one-dimensional power spectra are shown in
142 Chapter 4: Stochastic model for radar interferometry

fig. 4.32A, assuming isotropy. The diagonal lines in fig. 4.32A indicate the slope of
–5/3 and –8/3 power-law functions. The data show that all atmospheric situations
exhibit similar power-law behavior, with slopes varying between –5/3 and –8/3 for
the different scaling regimes. There is a range in the absolute power, indicated by
the vertical position of the spectra in the graph, of almost two orders of magnitude,
indicating more or less severe weather conditions. The upper line is corresponding
with the first interferogram in fig. 4.31, which indeed exhibits the most severe zenith
delay variability. The bold line is the average of all eight situations. In fig. 4.32B,
all power spectra have been multiplied by a factor, yielding the averaged power for
1 cycle/km. Comparison of the shape of the power spectra shows that the data are
not entirely scale-invariant, and three dominant regimes can be recognized, of which
two are atmospheric while the third one probably reflects noise.
Regime I, with exponent and fractal dimension see eqs. (4.1.14)
and (4.1.15), covers scales larger than the thickness of the turbulent layer. For these
scales, between and the size of the interferogram the ap-
proximation of two-dimensional turbulence can be applied, c.f. eq. (4.7.11). Over
these scales the characteristics of the total atmospheric column dominate. For ex-
ample, convective processes result in significant differences in the overall refractivity
between updrafts (warm and moist) and downdrafts (cold and relatively dry). Al-
though there is evidently atmospheric signal at scales larger than the interferogram
size, we can safely ignore this since it is outside the measurement capabilities of the
system. Stationarity of the physical process is always ensured due to the limitation
of the earth’s circumference, which enforces the flattening of the power spectrum for
large scales. Based on VLBI observations it is proposed that the slope of the power
spectrum is zero for scales larger than 3000 km (Treuhaft and Lanyi, 1987).
Regime II covers scales smaller than the thickness of the turbulent layer (the depth
 4.7 Atmospheric signal: turbulence 143

of the convective boundary layer, say 2 km), down to the resolution level. The
integrated refractivity has a much smoother behavior over these scales, indicated
by the steeper power exponent and the lower fractal dimension
The correspondence in the power slopes is clearly visible in fig. 4.32B, where all
curves overlap within this regime. The behavior in this regime has been observed in
several studies on SAR interferometry, see Goldstein (1995); Hanssen et al. (1998b);
Hanssen (1998); Ferretti et al. (1999a). Nevertheless, in contrast to some studies,
the combination with regime I indicates that the signal cannot be considered scale-
invariant for all spatial scales. In fact, assuming a –8/3 power-law for scales larger
than 5–10 km will increasingly underestimate the roughness and hence the variability
of the atmospheric signal.
Regime III is unlikely to have an atmospheric origin, since there is no physical
explanation for increased delay variation at such small scales. Instead, it is likely
caused by high wavenumber noise in the data. This can be a result of decorrelation
effects or, e.g., by interpolation errors when subtracting the reference DEM. In
these examples, regime III noise starts to influence the data for wavelengths less
than Although the driving mechanisms differ from regime I and II, it can
be advantageous to describe the statistical behavior of regime III the same way,
since the transitions between the regimes are gradual and the stochastic model can
be constructed uniformly. The power exponent in regime III approaches
and the lower fractal dimension indicating the very rough behavior at
these scales.
144 Chapter 4: Stochastic model for radar interferometry

Summarizing, the important observation from the power spectra is their similar
shape, even though the atmospheric circumstances were extremely variable, rang-
ing from calm, cold weather to severe storms with strong convection and humidity
gradients. These weather situations covered two orders of magnitude in the power
spectrum. Therefore, a reasonable approximation of the power spectra seems to be
based on the three regimes, using only the value (for 1 cycle/km wavenumbers)
for initialization, see table 4.5. This observation is used in the atmospheric model
proposed in section 4.7.4. The for the eight interferograms in fig. 4.31 are
listed in table 4.6.
Covariance function
Assuming second-order stationarity and isotropy. the covariance between two ob-
servations is a function of only the distance between the observations. Empirical
covariance functions C(r), where r reflects the distance between two observations,
are derived from the inverse DFT of the power spectra, see eq. (4.7.23), and shown in
fig. 4.33A. Clearly, the interferogram with the most severe atmospheric signal shows
the most dominant empirical covariance function. The variances (r = 0) are depen-
dent of the spatial support of the data, in this case the window size of 50 x 50 km
chosen for evaluation.
The standard approach to modeling the covariance function is by approximating
some parametric analytical form for C(r) which ensures positive-definiteness (Rum-
mel, 1990; Daniels and Cressie, 1999). However, since the power spectrum is non-
negative by definition, its Fourier transform is always a positive definite covariance
function (Bochner, 1959).
 4.7 Atmospheric signal: turbulence 145

Structure functions
The two-dimensional structure functions are derived from the eight interferograms.
From these a radial average yields the one-dimensional structure function, as shown
in fig. 4.33B. For comparison, the model proposed by Treuhaft and Lanyi (1987) is
plotted as the dash-dotted line. This model uses an effective height of the wet tropo-
sphere of 1000 m, and an initialization value derived from the standard deviation of
the zenith wet delay over a given time interval for sites at mid-latitudes (California,
Australia and Spain). It can be observed that these standard parameters provide a
reasonable model, but the variability in weather conditions over a coastal zone area
such as the Netherlands is too significant to allow for such a generalization. Hence,
the initialization parameters need to be adjusted to the specific weather conditions.

4.7.4 Empirical model

Based on the observations in interferograms, an analytical one-parameter model can

be constructed to describe the atmospheric signal. The application of atmospheric
models for radar interferometry is different from models used for, e.g., VLBI or GPS.
First, the interferometric data always reflect the superposition of two independent
atmospheric states. Second, the interferogram can be regarded as a snapshot—an
instantaneous acquisition of the integrated signal delay over the entire atmospheric
column—whereas the other methods measure variability in time at a fixed position.
These methods need to apply Taylor’s hypothesis of a “frozen atmosphere” to map
time series to spatial variation, using the wind speed (Taylor. 1938). Finally, the
spatial accuracy of a SAR interferogram is relatively high, say, 40 × 40 m, whereas
the atmospheric zenith delay for a technique such as GPS is a weighted average of
all slant delays to satellites at different directions and elevations using a mapping
146 Chapter 4: Stochastic model for radar interferometry

function. As a result, such a measurement cannot be regarded as a point measure-

ment anymore, but as an averaged value over an area with a diameter of several
kilometers, introducing aliasing effects. The InSAR data do not suffer from this
limitation.
Here we propose a new atmospheric model for radar interferometry, based on the
observed scaling regimes in section 4.7.3. The model assumes isotropy and no iono-
spheric influences. Furthermore, the influence of topography is ignored to obtain a
generic model. Topographic influence is discussed separately in section 4.8.
These results can be written as:

By defining f as the wavenumber in cycles/km and enforcing a

continuous function, we can initialize this model by one parameter, by measuring
the power spectral density at which yields P0 . The other two coefficients
are easily found to be and for

Figure 4.34A shows the empirical results of this approach, for eight independent in-
terferograms, suffering from atmospheric signal varying from thunderstorms to calm
weather. It is clear to see that the basic structure of all situations is similar, fol-
lowing the power-law behavior described above. The diagonal lines indicate –8/3
 4.7 Atmospheric signal: turbulence 147

power-law values. Note that it is not possible, as some authors have proposed, to
describe the scaling behavior with only one single power-law exponent. For wave-
lengths larger than 1.5 km none of the signals follow the –8/3 lines. Although
there is a strong similarity in shape, the difference between the absolute values of
these curves varies two orders of magnitude. It is this variation that is important to
consider while evaluating the quality of the interferometric data.
The covariance function derived from the model can be used to fill the variance-
covariance matrix based on the distance between each pair of pixels. The
diagonal elements consist of the model values and the phase variance derived from
the coherence of the data.
Simulation
Figure 4.35A shows a simulation of the model over a 50 × 50 km area. The transition
of regime I to regime II is chosen at a wavelength of 2 km, the transition of regime II
to regime III at a wavelength of 900 m. In fig. 4.35B an example of observed
tropospheric signal is shown for comparison. This situation corresponds with storms
during both acquisitions.
The isotropic scaling behavior as described above is applicable for many atmospheric
situations, from calm to stormy weather, without clouds to cumulonimbus clouds.
Although the scaling behavior is similar for all situations, only the absolute val-
ues of the variations are causing the differences in the delay of the wavefront. The
power initialization parameter can be observed over a relatively small area of the
interferogram, sufficient to estimate the entire behavior over larger scales. By trans-
forming the modeled spectra to the covariance functions, it is possible to estimate
the variance of the difference between points in the interferogram. This way, the
stochastic model for repeat-pass radar interferometry can be adjusted to account for
the atmospheric variability.

4.7.5 Conclusions

The stochastic modeling of atmospheric turbulent signal for radar interferometric

measurements is a key element in describing the quality of the data and the de-
rived parameters. Apart from the usual phase noise described by the interferometric
coherence, the variances and covariances introduced by the atmosphere need to be
included in this model as well. It has been shown that, although the state of the
atmosphere can be very different for various interferograms, the shape of the power
spectrum shows three distinct scaling regimes which are remarkably similar, inde-
pendent of the type of atmospheric behavior, as supported by Kolmogorov turbu-
lence theory, extended for ranges exceeding the 3D turbulence characteristics. The
absolute values of the observed power spectra differ two orders of magnitude.
Using the atmospheric model discussed in this section, it is possible to quantitatively
describe the covariance function of an interferogram using only one parameter. This
parameter can be estimated from a small patch of the interferogram, with flat terrain
and without deformation. This first order estimate may then be used to describe
covariances in the whole interferogram. An alternative approach is the estimation of
148 Chapter 4: Stochastic model for radar interferometry

the initialization parameter using, e.g., GPS time series in combination with wind
speed observations, see eq. (4.7.20). In this way, a single GPS receiver in the inter-
ferogram area would be sufficient to quantify the covariances in the interferometric
data.

4.8 Atmospheric signal: stratification

This section reports on the implications of vertical refractivity profiles in the tro-
posphere on the quality of Digital Elevation Models and surface deformation maps
derived from synthetic aperture radar (SAR) interferometry. A representative set
of 1460 radiosondes acquired over one year in the Netherlands is used to obtain
statistics for the differential delays between the two acquisition dates, and apply
these to simulate 1-day and k × 35 day intervals corresponding with ERS-1/2 orbit
characteristics. It is shown that differential delays can amount up to more than
1 cm for height intervals of 500 meters or more. For a 2 km height interval and
an interferometric baseline of 80 m such delays result in a height error of 180 m.
It is not possible to find a generally valid correction scheme for these delays using
surface meteorological measurements. Only in situ vertical profile measurements
such as radiosondes can be used to correct for these errors. To obtain a first order
indication of the extent of these effects on the accuracy of products derived from
radar interferometry, the rms of the delay is determined as a function of height. An
empirical expression for the rms is presented.
Introduction
Atmospheric stratification only considers variation of the refractivity along the ver-
tical. Assuming an infinite number of thin atmospheric layers, each with constant
refractivity, there will be no horizontal delay differences over flat terrain, even for
different refractivity profiles during both SAR acquisitions. This is due to the fact
that SAR interferograms are not sensitive to image-wide phase biases. However, for
hilly or mountainous terrain a difference in the vertical refractivity profile during
both acquisitions will affect the phase difference between two arbitrary resolution
cells with different topographic height, see fig 4.36A, and may cause an erroneous
interpretation. This effect has been recognized during deformation studies of mount
Etna by, e.g., Tarayre and Massonnet (1996); Massonnet and Feigl (1998); Delacourt
et al. (1998), and Ferretti et al. (1999a).
Since the resulting phase error in the phase difference between two resolution cells
has zero-mean (expectation value), and the observed phase gradients are often a
combination of topographic residuals, deformation, horizontal atmospheric hetero-
geneity and differential vertical stratification, it has not been possible yet to obtain
reasonable error estimates using interferometric data only. A simple and effective
way to address this uncertainty is to study vertical radiosonde profiles, and analyze
the statistics of the delay variation for every possible height interval. We analyzed
1460 radiosondes acquired 4 times daily from 1 Jan. 1998 to 31 Dec. 1998, located
in a moderate sea-climate at latitude 52.10° and longitude 5.18°. Although climate
conditions vary considerably over the globe, it will be shown that it is not the abso-
 4.8 Atmospheric signal: stratification 149

lute refractivity value, but the dispersion of the refractivity for a fixed height that
determines the amount of error for interferometry. Since water vapor is the main
factor influencing the variation of refractivity, we assume that the results reported
here can be regarded as representative for a large part of the globe, excluding polar
regions, where the atmosphere has a very stable stratification and the cold air cannot
hold much water vapor.
Theory
The geometric and delay terms for points p and q, see fig. 4.36, at different to-
pographic heights, during acquisitions and assuming zero-baseline, identical
incidence angles and no horizontal variation in refractivity N, can be written
as:

where is the geometric distance between point p and satellite s, projected onto
the vertical. The vertical delay between p and s at is indicated by
150 Chapter 4: Stochastic model for radar interferometry

The interferogram phases at points p and q are

and the contribution of the tropospheric delay to the interferometric phase difference
between point p and q is

From eq. (4.8.3) we conclude that whenever there will be a contribution

of the tropospheric vertical stratification in the interferogram. Note that this effect
will only affect points in the interferogram that have a different topographic height.
The total integrated effect is proportional to the height difference.
The delay is related to the refractivity profile by

Hence, we write the interferometric phase difference in (4.8.3) as

The difference in the vertical refractivity profile, eq. (4.8.5), can cause a significant
contribution in the observed phase difference between point p and q, see the results
presented below.
From eq. (4.8.4) it is clear that the dimensionless refractivity, integrated over unit dis-
tance, yields the fractional delay in parts per million (ppm). For example, N = 300
corresponds with a fractional delay of 0.3 mm/m. Therefore, we can regard inte-
grated refractivity values as cumulative delay. The two curves sketched in fig. 4.36B
indicate the cumulative delay at acquisitions and The delay differences be-
tween points p and q are and respectively. In the interferogram
the phase difference, expressed in delay will be

Hence, to be able to correct for the effect of vertical stratification, it is necessary to

know the difference in cumulative delay between and at point p and point q.
Results and discussion
Daily radiosonde launches at De Bilt, the Netherlands, measured height, pres-
sure, temperature, and relative humidity during their ascent. Using the Clausius-
Clapeyron equation the saturation water vapor pressure corresponding to the tem-
perature level is determined, which enables the conversion from relative humidity
 4.8 Atmospheric signal: stratification 151

to partial water vapor pressure (Stull, 1995). Using equations presented in (Smith
and Weintraub, 1953) we calculated the refractivity for every height level. For all
sondes acquired at noon during 1998, the results are shown in fig 4.37. Refractiv-
ity values are presented as fractional delay in ppm. Apart from seasonal variation
it can be observed that the profiles show a ragged behavior, indicating significant
variation in refractivity, especially at lower altitudes. Figure 4.38A shows the rms
of this variation as a function of height. After interpolation of the refractivity mea-
surements to a regular height interval we can perform the integration in eq. (4.8.5)
for every combination of two profiles during the year. Simulating different ERS-1/2
repeat intervals (1, 3 35, 70, 140 days), statistics of the contribution of atmospheric
delay due to vertical stratification can be obtained for all height intervals and time
intervals. Figure 4.38B shows the standard deviation of the zenith delay difference
between a point with zero altitude and points with altitudes up to 5 km. The size
of the standard deviation increases for longer time intervals, indicating that 1-day
intervals have better characteristics than intervals spanning seasonal changes.
With a precision better than 2–3 mm we find the following empirical model for
the standard deviation of the interferometric phase due to differential tropospheric
stratification:

where is the time interval in days, h represents height in meters, and

is a scale height. Above this height the variability of the refractivity is considered
152 Chapter 4: Stochastic model for radar interferometry

negligible. For the accuracy of a height difference one can simply use

Assuming a Gaussian distribution, these results imply that approximately 33% of the
interferometric combinations exhibit effects more severe than expressed in eq. (4.8.7).
For example, for a time interval of 175 days and a height interval of 2 km, 33% of the
interferograms will have more than one phase cycle error due to vertical tropospheric
stratification assuming ERS conditions. For a 100 m perpendicular baseline, this
translates to a height error of 100 m or worse. For a 1-day interval and 2 km height
interval it yields a height error of 76 m or worse.
Possibilities for correction
Since the size of the error described in the previous section is considerable, it is nec-
essary to investigate methodologies to correct for these errors. Three categories of
possibilities are investigated: vertical profile measurements, integrated refractivity
measurements, and surface measurements in connection with a model. It is obvious
that vertical refractivity profiles, acquired at the interferogram location during the
SAR acquisitions are the best option. In this case it is possible to insert the re-
fractivity values in eq. (4.8.5) and to determine the interferometric phase difference
for every pair of pixels. The second option, integrated refractivity or delay mea-
surements can result from, e.g., GPS observations. The problem with this type of
observation is that, in order to determine the integrated quantities in eqs. (4.8.4)
and (4.8.5), receivers are necessary in point p as well as in point q. Therefore, to
determine the interferometric phase component due to vertical stratification it is
necessary to have such a device at many (theoretically all) elevation levels, which
is impractical. The third option, surface observations, has been studied first by
 4.8 Atmospheric signal: stratification 153

Delacourt et al. (1998) on interferometric data of Mt. Etna and assumes that sur-
face observations of pressure, temperature, and relative humidity can be used to
approximate the vertical refractivity profile.
The radiosonde data used in this study are ideal to test the hypothesis of retrieving
a vertical refractivity profile from surface measurements since an unambiguous com-
parison between the error signal and the corrected signal can be performed. There-
fore, there is no risk of involving other parameters such as errors in the topographic
model or due to surface deformation in the results. We tested the Saastamoinen-
Baby model, which decomposes the total delay in a hydrostatic component and a wet
component (Saastamoinen, 1972; Baby et al., 1988). The hydrostatic component ap-
proximates the delay component based on surface pressure assuming an exponential
decrease with height. The derivative of this delay versus height yields the hydro-
static component of the refractivity. It is well known that this model is very accurate.
The wet component of the model is much harder to determine, since there is much
more variability with height. Baby et al. (1988) applied the coarse assumption that
the relative humidity at the surface remains constant over a certain height interval,
above which it reduces to zero. An expression for the wet component of the delay,
using two climate-dependent constants, surface relative humidity, and temperature
linearly decreasing with height can be found in Baby et al. (1988). Differentiating
this refractivity-height curve yields the wet component of the refractivity.
Summing the hydrostatic and the wet component of the refractivity yields the
“model”-refractivity which can be compared with the “true” refractivity
as obtained from the radiosonde. To study the feasibility to use this model
to correct the atmospheric signal due to vertical stratification in the interferogram,
we simulated the interferometric error signal by subtracting the refractivity profiles
during two days, both for the true as well as the model refractivity, and subtracting
both results:

which yields the residual refractivity for that simulated interferogram. Com-
parison of the cumulative sum of and denoted and
indicates whether the error signal due to vertical stratification is reduced or not. If
for every height level

the correction improves the results. Applying this scheme to all interferometric com-
binations with yielded, with increasing height, 0–50% significant
improvement (correction of errors by 4 mm or better). In 0–35% of all simulations,
the error signal increased after correction using the surface observations.
The unpredictable behavior of the wet component of the refractivity is the main
reason for these disappointing results. In many cases, surface humidity measures
do not reflect the humidity of the total profile. An obvious example of such a
situation is fog, where surface relative humidity is 100%, which does not necessarily
imply high humidities at higher levels. From the analyzed data, it appears that
154 Chapter 4: Stochastic model for radar interferometry

such misinterpretations are common. The fact that interferometric data reflect a
linear combination of two different profiles makes it nearly impossible to model the
behavior of vertical stratification based on surface observations.
Conclusions
The accuracy of DEMs obtained from repeat-pass radar interferometry is signifi-
cantly influenced by atmospheric signal due to turbulent mixing and vertical strati-
fication. An empirical relation between the size of the error due to the latter com-
ponent has been derived for different height intervals and temporal baselines, using
a set of radiosonde observations acquired during 1998. The analyzed dataset is re-
stricted to one location on earth, which limits conclusions on global scales. However,
since not the total refractivity but the variability of refractivity influences the signal
in radar interferograms, it is expected that the results obtained here give a correct
order of magnitude of the error.
Correction of the error due to vertical stratification is only possible using vertical
profile measurements. Surface observations combined with a tropospheric model are
in general unreliable, whereas integrated refractivity observations such as obtained
using GPS are inadequate for correction.

4.9 Error propagation and joint stochastic model

In this chapter the most important error sources for repeat-pass radar interferometry
have been analyzed. The joint stochastic model for a specific radar interferogram
can be constructed by combining all these models in the variance-covariance matrix
see eqs. (4.2.1) and (4.2.2). The error statistics affecting the data can be
categorized in two classes: single-point statistics and multiple-point statistics. The
single-point statistics describe the quality of a single observation (resolution cell)
in the image, and are described by the PDF for that observation. It is assumed
that the observations are not correlated. Multiple-point statistics consider possible
correlations between the observed values at resolution cells with different locations,
and are described using the covariance function, assuming second-order stationarity.
The phase variance can be (i) predicted a priori or (ii) estimated from the data.
For a priori prediction, the characteristics of the interferometric geometry, the sensor,
and the processing are used together with assumptions on the terrain to obtain
predicted coherence values. The product of the coherence values and the number of
effective looks are then converted to phase variance, using eqs. (4.2.24) and (4.2.27),
see fig. 4.39. Nevertheless, this approach can be valuable only to estimate an upper
bound for coherence and a lower bound for phase variance, since the influence of
temporal decorrelation is often spatially variable and difficult to assess. For rocky
surfaces, such as lava flows, temporal decorrelation will be minimal even over years,
whereas decorrelation due to anthropogenic activity can happen overnight. Usually,
knowledge of terrain vegetation in combination with the wavelength of the sensor and
some experience can be used to obtain a coarse qualitative indication of the effect
of temporal decorrelation. Regarding the geometric decorrelation it is stressed
that these equations hold for distributed scattering mechanisms. Corner reflector
4.9 Error propagation and joint stochastic model 155
156 Chapter 4: Stochastic model for radar interferometry

type point scatterers (artificial or natural) can remain coherent over a wide range of
baselines. For these targets, the measure of coherence is generally unsuitable due to
the ergodicity assumptions. Thermal decorrelation is dependent on the radar cross-
section of the reflections. For example, an airport runway will have an extremely
low radar cross section, whereas mountains with a slope identical to the look
angle will have an extremely high value for In general, however, this is difficult
to predict.
The advantage of a priori prediction of the phase variance is that some information
is available even before the data are ordered. This way, suitable interferometric pairs
can be chosen from the database. The disadvantage is that usually the quality of an
interferogram is spatially variable and temporal decorrelation is often very difficult
to predict quantitatively. The alternative is (a posteriori) phase variance estimation
from the processed interferogram. In this case, the coherence is simply estimated
using an estimation window, see eq. (4.3.2).
The phase variances (estimated or predicted) can be used to fill the diagonal of the
variance-covariance matrix as

Spatial correlation in the interferogram data can occur due to orbit errors and at-
mospheric signal, as described in sections 4.6 and 4.7. Since orbit errors result
in a long-wavelength trend over the entire interferogram, second-order stationarity
will not be reached. The same holds for phase ramps due to atmospheric influ-
ences. Although second-order stationarity is not necessary for quality assessment
using the structure function, it is necessary when using the covariance function for
data adjustment and filtering. Since the covariance function is needed to fill the
variance-covariance matrix, the data need to be detrended, which removes the long
wavelength orbit errors.
The remaining spatial correlation, mainly due to atmospheric signal in the data, is
described using the atmospheric model proposed in section 4.7.4. The initialization
value can be estimated from (parts of) the data or from other sources such as
GPS time series. If no information is available, a conservative value can be chosen
for initialization. Using eq. (4.7.23), the positive definite covariance function is
obtained.
The variance-covariance matrix of the atmospheric delay, cf. eq. (3.1.11), can
be constructed using the horizontal distance between every pair of observations and
the covariance function. The covariance corresponding to this distance is inserted
at the appropriate position in yielding a symmetric variance-covariance ma-
trix, as shown in fig. 4.40. The first column of this matrix corresponds with all
combinations of observations with observation at the upper left position in the
original interferogram. Since the vectorization of the observations is performed as in
eq. (3.1.1), see fig. 3.1, the covariance behavior for observations in the same column
 4.9 Error propagation and joint stochastic model 157

of the interferogram is smooth, whereas the step to the next column of the inter-
ferogram is a discrete transition. Performing this procedure for all combinations
with all observations, we find as in fig. 4.40. Note that for this example, a very
small interferogram of 12 × 12 pixels was simulated, with a posting of approximately
160 m. Notice that the number of elements in the increases quadratically
with the interferogram size.
The structure of is a Block Toeplitz matrix with Toeplitz blocks, also known as
a BTTB matrix. In fig. 4.40, matrix is a 12 × 12 block matrix, each block being
a 12 × 12 Toeplitz matrix (Strohmer, 1997). For BTTB matrices, fast inversion
techniques exist, see, e.g., Kailath et al. (1978); Bitmead and Anderson (1980);
Strang (1986); Ammar and Gragg (1988); Brent (1991), and Strohmer (1997).
The joint stochastic model combines the influence of the single-point statistics (point
variances) and multiple-point statistics (covariance functions). By adding and
we find the joint variance-covariance matrix C:

Simply adding the variance in to the diagonal of is possible due to the distinct
difference in the driving mechanisms of these contributions, which ensures that they
are uncorrelated.
It is obvious that the inversion of the variance-covariance matrix C is a significant
numerical challenge, especially when using an entire single-look interferogram (more
158 Chapter 4: Stochastic model for radar interferometry

than 100 million observations). However, with we can decompose the

inversion of C as (Golub and Van Loan, 1996)

Hence, finding an analytic expression for the inverse of BTTB matrix based on
the atmospheric model and the pixel locations could aid a relatively fast inversion
of the joint variance-covariance matrix.
Consideration of the influence of vertical atmospheric stratification, as described in
section 4.8, is only possible when initial topographic heights are known or derived
from the data recursively. If such information is available, the covariance between
two points in the image at a different height level can be added to
For completeness, we remark that we will leave the problem of the integer phase
ambiguities outside the scope of this discussion. For now, we assume that phase
unwrapping does not add significantly to the total error budget.

4.9.1 Height error

The propagation of an error in the interferometric phase to topographic height differ-

ences has been derived by, e.g., Rodriguez and Martin (1992); Zebker and Villasenor
(1992); Zebker et al. (1994a). In these approaches the variance of the topographic
height at point p follows from eq. (2.4.12)

where the phase variance is found as a function of the coherence and the multilook
number using eqs. (4.2.24) and (4.2.27), assuming distributed scattering. This ex-
pression, however, suggests an absolute point positioning accuracy, a property which
is not justified since the obtained height observations are relative to a reference sur-
face. In practice, the relation with the reference surface can only be obtained by
examining spatial gradients between resolution cells. This concept implies that, in
order to obtain the variance of the topographic height, an arbitrary reference pixel is
needed and the height is in fact a height difference between that reference pixel and
the pixel under consideration. As a result, the spatial correlations in the data, e.g.,
as introduced by the atmosphere, need to be included, which yields the variance of
the height difference between point p and point q:

where short hand notation is used. In this expression, two neighboring

resolution cells will have a small since the covariance function of the phase
has a high value, as the atmospheric contribution for both resolution cells are nearly
 4.10 Conclusions 159

identical. For two resolution cells with a large spatial separation, the variance of the
height difference will be large, since the covariance function of the phase has a
small value, as the atmospheric contribution for both resolution cells will be almost
uncorrelated.

4.10 Conclusions

This chapter gave an overview of the error sources in repeat-pass spaceborne radar
interferometry. It summarized how the error sources can be described mathemati-
cally and how a stochastic model can be devised which includes these influences. It
is proposed to include the spatial variation of the atmospheric signal as a covariance
function in the stochastic model. In an interferogram, the functional model relates
the observations (phase values for every pixel) to the unknown parameters (e.g.,
the parameters of a deformation model). The stochastic model appoints a variance
for every single pixel and covariances between any combination of two pixels. The
determination of the variance values has been discussed in sections 4.1–4.4. Covari-
ance functions, mainly considering atmospheric signal, were modeled and reported
in section 4.7-4.8. Although the formulation of the stochastic model is a step for-
ward in the analysis and interpretation of interferometric data, there are numerical
challenges to overcome, e.g., the inversion of the BTTB variance-covariance matrix.
 Chapter 5
Data analysis and interpretation for
deformation monitoring

This chapter focuses on the application of repeat-pass (differential) interferom-

etry for deformation monitoring. Some case studies are presented which serve
as an example for the feasibility and limitations of the technique.
key words: Deformation, Differential interferometry.

Deformations of the earth’s surface occur at many different scales and with vary-
ing magnitudes. To be observed and quantified by repeat-pass spaceborne radar
interferometry the process needs to fulfill certain conditions. In terms of scale, the
dominant scales should be less than the size of the interferogram. Generally, a small
discrete deformation between two neighboring resolution cells can be more easily reli-
ably quantified than smooth spatial variations which might mimic, e.g., atmospheric
signal or trends due to orbit inaccuracies. In terms of magnitudes, the deformation
signal needs to be at least, say, two times as much as the noise source related to
that specific scale. For example, for neighboring resolution cells, the atmospheric in-
duced noise can be safely ignored, and the deformation signal only needs to be more
than the phase noise due to decorrelation. On the other hand, for deformations
spanning a large area, atmospheric noise can be significant, forcing larger deforma-
tions to take place before obtaining a satisfying signal-to-noise ratio. In practice,
e.g., for continuous-type deformation processes which span large time intervals, this
might imply that only interferograms with a long temporal baseline can be used.
Many of the deformation studies using interferometry have focused on phenomena
that fulfilled both the scale and the magnitude (SNR) requirements, and achieved
spectacular, uniquely interpretable results.
In this chapter, some case studies will be presented which have three common dif-
ficulties compared with the more opportunistic studies mentioned above. First, the
problem of temporal decorrelation will result in isolated patches of reasonable co-
herence, surrounded by areas which are completely decorrelated. Sometimes these
coherent patches can be only one or a few resolution cells wide. This poses an identi-
fication problem: it needs to be decided whether a patch is reliable (coherent) or not.
The conventional solution for this problem is by estimating the coherence of the im-
age and thresholding on some predefined coherence value. This approach, however,
162 Chapter 5: Data analysis and interpretation for deformation monitoring

is rather pragmatic and completely fails in the case of only one or a few reliable pix-
els. This is due to the fact that a relatively large coherence estimation window needs
to be used to obtain an unbiased coherence estimate, and the ergodicity assumption
is not satisfied for a single coherent observation.
Second, after successful identification of the patches (pixels) with a reliable and
accurate phase value, the patch needs to be unwrapped. Although this might be
possible within the patch, difficulties arise when separated patches need to be con-
nected. Especially when the distance between those patches is too large it appears
to be almost impossible to connect them. The only way to solve this problem is to
make the problem less under-determined, that is, add more independent observa-
tions (other interferograms) or introduce a priori information on the characteristics
of the problem.
Finally, the problem of connecting the identified patches is even more complicated by
the influence of atmospheric signal, which might result in a bias of an entire coherent
patch. The combination of this problem with the general problem of connecting the
patches can be another reason to combine many interferograms, aiding interpretation
and analysis.
In the following we start with a short discussion on the decomposition of the three-
dimensional displacement vector in terms of the radar range-change observation. To
obtain some experience with the accuracies of the deformation measurements on
a pixel-by-pixel base, we performed an experiment using a set of corner reflectors,
which were intentionally displaced in height during the ERS-1/2 “tandem mission.”
These results are presented in section 5.2. In the same area, with varying land-use
due to agricultural activities, it is investigated how conventional coherence estima-
tion windows would need to be applied to obtain a coherent signal over a time span
of 3.5 years. This is reported in section 5.3. A more significant occurrence of land
subsidence due to the activities of a geothermal plant in Mexico is discussed in sec-
tion 5.4, and a concise presentation of the extended and complicated deformation
pattern associated with the 17 August 1999 earthquake in Izmit, Turkey is given in
section 5.5.

5.1 Decomposition of the displacement vector

As a radar is only capable of measuring path length differences in its line of sight
or slant-range direction, deformation interferograms imply a non-uniqueness and
need to be interpreted with care. A three-dimensional displacement vector with
components and in North, East, and Up direction respectively, will be
projected to one slant-range component in the radar line-of-sight (LOS). For a
satellite orbit with heading (azimuth) see fig. 5.1A and 5.1B, we find

where corresponds with the angle to the azimuth look direction, which
is perpendicular to the satellite heading, for a right-looking satellite. The inci-
dence angle is denoted by For an incidence angle of and a heading
we find a sensitivity decomposition of a slant-range deformation of
 5.2 Corner reflector experiments 163

[0.92, –0.07,0.38] Therefore, assuming only vertical deformation, one

fringe corresponds with about 26 mm.
Due to the projection of the three-dimensional displacement vector to the radar
line-of-sight, it is not possible to retrieve the full displacement vector. Using a
combination of an ascending and a descending interferogram, two of the three com-
ponents can be retrieved. For the third component, assumptions are necessary on
the characteristics of the displacement. For example, for glacier flow it is often as-
sumed that the flow is parallel to the surface, which confines the vector along the
gradient of the elevation model (Mohr, 1997).

5.2 Corner reflector experiments

Parts of the northeastern part of the Netherlands (Groningen, see fig. 5.2) have
been exploited since 1964 due to their large natural gas reservoirs. Due to this gas
extraction, a maximum land subsidence of 23 cm has occurred between 1964 and
1998 in the center of a bowl shaped subsidence pattern (NAM, 2000). The short-
est distance from the center of the subsidence bowl to a presumably stable point is
approximately 23 km. This implies a maximum subsidence rate of ~7 mm/yr and
a subsidence gradient of ~0.3 mm/km/yr. It is expected that the total amount of
subsidence at the end of the production period will be around 38 cm in 2050 (NAM,
2000). Assuming a comparable spatial extent of the general subsidence pattern,
this results in an expected maximum subsidence rate of ~3 mm/yr and a subsi-
dence gradient of ~0.1 mm/km/yr. Extended leveling campaigns, surveyed every
five years, produced a robust network of benchmarks, which is now regularly sur-
veyed using GPS receivers as well. However, based on current experience, the GPS
164 Chapter 5: Data analysis and interpretation for deformation monitoring

network results are not dense and accurate enough to successfully replace leveling
(NAM, 2000). Based on the presumed high spatial density of SAR acquisitions it
was proposed to investigate the feasibility of radar interferometry for this specific
subsidence problem.
An extensive radar interferometry experiment is performed within the bounds of
the subsiding area, involving time series with a long temporal baseline, tandem
acquisitions for monitoring the magnitudes of the atmospheric signal, corner reflec-
tor deployment as stable reference points, and GPS observations to investigate the
possibility of estimation or correction of atmospheric signal from the data.
Since the Groningen landscape contains mainly agricultural areas it suffers from
severe temporal decorrelation, which makes interferograms decorrelate over more
than a couple of months. For some areas the decorrelation can be instantaneous due
to anthropogenic activities. The topography of Groningen is very flat, over large
areas less than a few meters of height variation.
In this section, we report on an experiment to assess the accuracy of the technique in
terms of the vertical movement of the corner reflectors between two SAR acquisitions.
The following section will comment on the possibilities for finding stable points for
accurate phase estimates. The influence of the atmosphere over this region has been
discussed in section 4.7.
 5.2 Corner reflector experiments 165

5.2.1 Introduction

During the Groningen Interferometric SAR Experiment (GISARE), a test was per-
formed in detecting vertical corner reflector movement in between two ERS tandem
passes of 16 and 17 March 1996, similar to an experiment performed by Hartl et al.
(1993), Hartl and Xia (1993), and Prati et al. (1993). Eight corner reflectors were
installed in the region. To allow accurate monitoring of the heights, all are located
within less than 100 m distance of a fixed benchmark of the vertical reference da-
tum (NAP). Of the eight installed corner reflectors, two were moved upward using a
screw mechanism. In this section it is analyzed if this movement is detectable in the
interferogram, how accurate a height change of a corner reflector can be monitored
using differential SAR interferometry, and which parameters are of influence for the
statistics of the experiment.
Theory
Following the notation conventions of section 2.4, the received phase between
satellite 1 and resolution cell P is the summation of two independent components, a
scattering part and a path length part Assuming that the scattering
part remains relatively unchanged between the two acquisitions, phase differences
are caused only by path length differences, hence
Path length differences arise due to (i) the oblique viewing geometry in
combination with the baseline, (ii) topographic height variation within the scene,
(iii) local surface deformation, (iv) spatial atmospheric delay differences, and (v)
spatial differences in the moisture or, more general, dielectric constant.
Phase ramps caused by the baseline geometry and earth curvature are reduced using
precise orbit information (Scharroo and Visser, 1998). For spatial scales associated
with the corner reflector, residual phase ramps can be removed safely. The topo-
graphic effect can be estimated using the height ambiguity, see eq. (2.4.14), p. 37.
For the interferogram under consideration, the perpendicular baseline
which results in a height ambiguity of approximately 516 m. Since the maximum
height differences in the GISARE test area do not exceed 20 meters, topographic
phase differences of 1.4 degrees can be expected. These differences are negligible
with respect to the thermal noise in the phase measurements. Moreover, the height
of the corner reflector with respect to its surroundings is ~1 m, which results in a
topographic phase difference of 0.7°, or an error in the vertical deformation mea-
surement of This effect is beyond the measurement accuracy and can
be ignored.
Phase differences due to lateral atmospheric delay differences within the scene will
not effect the difference between two neighboring resolution cells, see section 4.7.
Moreover, examining the total interferogram of this scene, see fig. 5.3, it can be
observed that atmospheric inhomogeneities are extremely limited, apart from some
slight atmospheric gravity waves. As a result, the only phase variations over small
spatial scales around the corner reflectors need to be due to relative deformation
between one resolution cell and another, although the unknown behavior of the
dielectric constant might have some limited influence as well.
166 Chapter 5: Data analysis and interpretation for deformation monitoring

5.2.2 Ground truth

At March 15, all eight corner reflectors were leveled with respect to a fixed reference
point, usually a concrete GPS or NAP marker in the vicinity of the reflector. In
table 5.1 this leveling is denoted with L1. ERS-1 acquired its radar data on March
16, at 11:31 LT. After this pass, two selected reflectors, 1 and 5, were raised approx-
imately 1 cm. To determine this vertical movement as accurately as possible, before
and after the movement the reflector was leveled with respect to a local, temporary
benchmark. These surveys are denoted by L2 and L3, respectively. Exactly 24 hours
after the pass of ERS-1, ERS-2 acquired its data of this scene from nearly the same
orbit as ERS-1. Finally, one day later the first leveling was repeated with respect
to the GPS or NAP marker, here denoted by L4. In table 5.1 the chronological
sequence of events for reflector 1 and 5 is listed. As the corner height adjustment
was applied shortly after the pass of ERS-1, the result of leveling L2 and L3 can be
considered very reliable. However, the time interval between the corner adjustment
and the pass of ERS-2 is approximately 24 hours. Therefore, we need to test if
 5.2 Corner reflector experiments 167

between these times any additional corner movement took place, perhaps due to an
unknown disturbance. To study this possibility we examine the difference

The left-hand part is referred to as the long epoch, the right-hand part as the short
epoch.
The leveling surveys were performed using a closed loop, which yields two height
differences h1 and h2 between the reference point and the corner reflector. This
procedure gives a conventional mean and standard deviation a defined by

The mean height difference between the before and after measurements is now
defined using the difference of the means

and standard deviation

The results of the measurements for reflectors 1 and 5 are presented in table 5.2, and
graphically in fig. 5.4. Based on the curves for L3-L2 (the short epoch) and L4-
L1 (the long epoch), indicated by the dashed lines in fig. 5.4, two conclusions are
possible: (i) Between L3 and L4, which is two days later, reflector 1 has moved
downward 0.31 mm and reflector 5 has moved upward 0.67 mm, or, (ii) there has
not been any additional movement, apart from the movement we induced ourselves.
In the latter case, the differences between the mean values of the two survey combi-
nations is caused by measuring noise. Since the differences are extremely small, it is
assumed that the latter possibility is correct. In this scenario the height difference
measurements of the short and the long epoch are considered to be uncorrelated, and
standard error propagation yields the average reflector movement with associated
standard deviations. These values are shown in table 5.2 and indicated by the solid
line in fig. 5.4.
168 Chapter 5: Data analysis and interpretation for deformation monitoring

5.2.3 InSAR deformation measurements

To obtain an estimate of the phase difference between the corner reflector and the
surrounding area, the phase of the corner reflector is determined at sub-pixel accu-
racy using complex interpolation. The phase of the surrounding area is called the
reference phase. The phase difference between corner reflector and reference phase
can be mapped to a change in the elevation of the reflector. Although only two
of the reflectors are manually adjusted in elevation, any natural movement, which
might occur between the two acquisitions, might be detected by the leveling surveys
as well. Therefore, we analyze the observed deformations of all eight reflectors.
The corner reflectors can be manually detected in the interferogram amplitude using
a topographic map as a reference. An area of 256 × 256 complex pixels around each
reflector is stored for further evaluation. Using the fringe frequency parameters
of the interferogram, a flat-earth phase is removed in each of these areas. In the
complex image, an area of 32 × 32 pixels centered around the corner reflector is
interpolated to 1/16 of a pixel using spectral zero-padding at the spectral minimum.
The interferometric phase of the corner reflector is calculated at the maximum of
the peak power in this interpolated image. Besides the peak power and peak phase
also the peak location and 3dB width are stored.
The reference area consists of four sub-distributions of 80 × 80 m within an area of
1 × 1 km centered at the corner reflectors position, see fig. 5.5. To avoid contami-
nation by the sidelobes of the corner reflector, a strip of 11 pixels in range and 50
lines in azimuth (approximately 200 meters wide) is excluded from reference phase
estimation. Due to these two perpendicular strips, four rectangular areas of approx-
imately 400 × 400 meters (20 × 103 pixels) are created around the corner reflector.
Within each of these areas, the sub-distributions with minimal phase variance are
selected. The four sub-distributions contain 80 phase measurements (4 pixels in
 5.2 Corner reflector experiments 169

range and 20 lines in azimuth), corresponding with an area of 80 × 80 meters. The

mean phase value of the four sub-distributions is chosen to be representative for the
background of the corner reflector. The standard deviation of the reference phase is
derived using plain error propagation.
Phase statistics of the corner reflector
Since the complex value of a SAR pixel is the coherent superposition of all scatterers
in the corresponding resolution cell, the interpolated value of the interferometric
phase of the corner reflector is mainly disturbed by the influence of the surrounding
scatterers within the resolution cell, as shown in fig. 5.6. Using an estimation of the
power of the background reflections the radius of the corresponding error circle
is determined by

The expectation of the background power is determined in the master as well as

in the slave image, using a mean value of four areas of 2 × 2 pixels around each corner
reflector. The corner reflector power is also determined in master and slave image
using complex interpolation, see table 5.3. Average values for the standard deviation
of the phase are used in the evaluation of the corner reflector movement.

5.2.4 Results

The results of the experiment are listed in table 5.4. The first part of the table shows
the observed phase values for the background and the corner reflectors It is
170 Chapter 5: Data analysis and interpretation for deformation monitoring

assumed that the phase difference between these values is less than although
the is important to be considered in the interpretation of the results.
Using the local incidence angle and the wavelength the range displacement
and vertical displacement is obtained.
The second part of the table shows the accuracies of the interferometric observations.
The standard deviations of the reference phase value results from the averaging
of all reference phase observations. The standard deviation for the reflector phase
was listed in table 5.3. Using the same mapping with the local incidence angles,
these values are mapped to the vertical height values and and consequently
to the standard deviation of the vertical deformation of the corner reflector
The third part of the table lists the height changes observed by the leveling
surveys, and their associated standard deviations Finally, the last part of
the table shows the differences between the leveling and the InSAR results.
5.2 Corner reflector experiments 171
172 Chapter 5: Data analysis and interpretation for deformation monitoring

5.2.5 Discussion

The results listed in table 5.4 are visualized in fig. 5.7. The dots represent the
movement as observed by InSAR and the leveling results. Error bars delimit the
interval. It can be concluded that the movement of five of the eight reflec-
tors is correct within the 95% probability curve. Within this group, the movement
of reflector 1 is captured almost perfectly by the SAR analysis. For two of the
reflectors the difference with the leveling results is between 3 and 5 mm, whereas
and for one of the two intentionally elevated ones the difference is
almost 12 mm (with It is remarkable that the latter observation
has a which is twice the accuracy of the other corner reflectors. This is an
indication that changes in the reference area (the earth’s surface surrounding the
reflector) between the two SAR acquisitions might be the cause of this bias. In-
deed, the surface around reflector 5 was a bare and wet ploughed field. It can be
suggested that changes in moisture (dielectric constant) of this field have resulted
in a coherent phase change during the one-day time interval. That such changes
can occur very locally is shown, e.g., in fig. 5.8, where local coherent phase changes
related to agricultural fields are observed. The temperatures during both acqui-
sitions were just above the freezing level, 1°–3°C. Therefore, it is very likely that
during the night between the acquisitions the surface of the fields were frozen. If
 5.3 Groningen coherence estimation 173

the ploughed field surrounding reflector 5 reacts differently to freezing and thawing
compared with the smooth grass land surrounding reflector 1, it can be expected
that phase changes relate either to the deformation (swelling/deflation) of the field,
or to different moisture content.

5.3 Groningen coherence estimation

As discussed in the introduction of this chapter, the identification of coherent pixels

forms one of the most severe problems in areas that suffer from significant temporal
decorrelation. In fig. 5.9A it is shown how the coherence over the area of Groningen
174 Chapter 5: Data analysis and interpretation for deformation monitoring

has decreased over a time interval of 3.5 years. From this coherence image it may
seem that no significant information can be retrieved in this interval, since the overall
tone of the image has a constant and low coherence value. On the other hand one
might conclude that this average coherence value (~0.3) is still high enough to be
interpretable in terms of a reliable interferometric phase. After closer analysis of the
data, see fig. 5.9B, isolated points with a relatively high coherence estimation can
be detected, especially over the urban areas.

As discussed in section 4.3, the discrimination between areas of different coherence is

mainly influenced by the number of independent samples in the coherence estimation
window. If this number is too small, low coherence areas become biased towards a
higher coherence estimation. Figure 5.9A, for which an estimation window of 2 × 11
samples was used, shows this effect. Although almost all coherence in this (mostly
agricultural) area has disappeared, a bias of ~0.3 is apparent in the data. Therefore
discrimination between different coherence levels becomes difficult.

In fig. 5.9C it is shown how an increase of the coherence estimation window from
2 × 11 to 20 × 110 samples yields an improved discrimination between areas of low
coherence, which is due to the corresponding drop in coherence bias. In the figure,
this is obvious from the red tone of the decorrelated areas. Note that although a
continuous color wheel proved to be the best visualization for these data, coherence
levels above 0.95, which are also colored red, can be neglected with this estimation
window. On the other hand, areas with a relatively high coherence can be detected
as the yellow and green spots. Obviously, the larger estimation window implies a
decrease in the absolute coherence levels with respect to those of small window sizes.
Therefore, the range of coherence values in the image reduces, whereas the reliability
of the estimates increases.

Due to the better possibilities to discriminate between coherence levels, thresholding

at a chosen low level becomes possible when applying large coherence estimation win-
dows. Figure 5.9D shows how the interferometric phase values corresponding with
pixels with a coherence higher than 0.1 still reveal interpretable phase information.
In this masked interferogram, a multilook window equal to the coherence estimation
window (20 × 110) was applied. Although disturbing influences of the satellite orbits
and atmospheric heterogeneities still form a major restriction for an unambiguous
interpretation, this method of phase analysis reveals promising prospects for the
study of interferograms spanning long time intervals and severely decorrelated inter-
ferograms. The coherence preserving areas are in this case mainly of anthropogenic
origin.

A significant limitation of using large coherence estimation windows is the sensitiv-

ity for non-coherent pixels. When decorrelated pixels have a significant magnitude,
perhaps due to speckle effects, they will contaminate the phase values of the co-
herent pixels. This problem can be circumvented by using many coregistered SAR
images and forming a temporal instead of a spatial coherence estimation window, as
demonstrated by Ferretti et al. (2000).
 5.4 Cerro Prieto geothermal field 175

5.4 Cerro Prieto geothermal field

The Cerro Prieto Geothermal Field (CPGF) is located 30 km south-east of Mexicali,

Baja California, Mexico, see fig. 5.10. Centered between the right-lateral, strike-slip
176 Chapter 5: Data analysis and interpretation for deformation monitoring

Imperial fault and Cerro Prieto fault, the area is a pull-apart basin in the southern
part of the Salton Trough. The extraction of water and steam from the geothermal
system, at a depth of 1500–3000 m causes a maximum subsidence rate in the order of
8 cm/yr (Fabriol and Glowacka, 1997). Interferometric SAR observations, obtained
between May 1995 and August 1997, are used to evaluate the potential of the tech-
nique for monitoring the subsidence. The spatial characteristics of the deformation
are analyzed using linear combinations of differential interferograms. Information
on deformation, production, and recharging by fluid injection is used for a coarse
geophysical interpretation of the deformation mechanisms.

5.4.1 Background

The Salton Trough and Gulf of California are the results of tectonic activity that
created a series of spreading centers and transform faults, linking the East-Pacific
Rise with the San Andreas fault system. (Pacific–North-American plate boundary).
As a result, at Cerro Prieto, the earth’s crust is being pulled open (“pull-apart
 5.4 Cerro Prieto geothermal field 177

basin”) an thinning. An upwelling of magma from the asthenosphere, combined with

intrusions in the form of dikes and sills is the main heat source for the reservoir.
This is causing groundwater in sandstones and gray shales from the Colorado River
delta to be heated at depth. Formation temperatures at 2500 m are estimated
at 350°C. One of the most notable manifestations of the thermal activity in the
area is the Cerro Prieto volcano, with an elevation of about 210 m. Since the
Mexicali Valley floor was a part of the Colorado river delta, it is a very flat area
with agricultural activity. The area west of CPGF is dry and shows geothermal
surface manifestations such as mud volcanoes and fumaroles. The climate in the
area is extreme: temperatures vary from –2°C in winter to +47°C in summer.
Rainfall is very low, with an annual mean of 80 mm.
The CPGF, operated by the Mexican Comisión Federal de Electricidad (CFE), is
one of the largest producing geothermal fields in the world. It is exploited since
1973 and currently involves three power plants with more than 100 operating wells,
covering Since 1989, the natural fluid recharge of the reservoir has been
supplement by (artificial) injection in 7 wells. In October 1996, the three Cerro
Prieto power plants, see fig. 5.11B, were generating about 560 MW of energy based
on a total production (steam and water) of approximately 12.000 T/h (tons per
hour or During that month, some 2.300 T/h of cooled geothermal brine were
being reinjected into the reservoir (19% of the fluids produced). Over the years, this
substantial net extraction of mass has resulted in systematic land subsidence.
The Cerro Prieto wells produce a mixture of liquid, vapor, and gases. As the fluid
from the reservoir rises to the surface it flashes in the wellbore. The steam and
gases are sent to the power plant turbines. The separated brine is redirected to
the evaporation pond, which is clearly visible in the radar images. Part of the
brine is reinjected in the injection wells, see fig. 5.11B. The subdivision in the three
well groups CP-I, CP-II, and CP-III is caused by the different reservoir and well
characteristics. Every well group has a separate power plant. The average distance
between the wells is about 400 m.

5.4.2 Leveling results

Ground subsidence has already been observed from first-order leveling between 1977
and 1979. However, due to a strong earthquake in the region in June 1980, it was
not clear how much of the subsidence in these and later surveys was due to fluid
extraction. In general, leveling and gravimetry show a subsidence rate in the order
of 6 cm/yr, averaged over the last 35 years, with a maximum rate of 10 cm/yr.
In order to validate the results of the time period covered by the SAR interferograms,
results from leveling campaigns performed by CFE in 1994, 1996, and 1997 have been
used. The leveling network of 1996 covered only a limited part of the 1994 and 1997
surveys. The interpolated results for the 1994–1997 epoch are shown in fig. 5.12.
The subsidence is imaged using a 10 cm color interval, to provide more sensitivity
for gradient differences. The general subsidence bowl is located at the center of
CPGF and shows a subsidence rate of ~30 cm over the three years (~10 cm/yr).
The iso-line corresponding with zero-subsidence is indicated by the striped-dotted
178 Chapter 5: Data analysis and interpretation for deformation monitoring

line in the lower left of the area, at a distance of approximately 9 km of the main
subsidence bowl. The iso-line of – 25 cm subsidence corresponds approximately with
the outer margins of the three well groups.
A small pit with a maximum subsidence of 55.7 cm (~18 cm/yr) and a diameter
of 1.5 km is located between the railroad and the evaporation pond. Although a
limited number of benchmarks contributes to this pit, the interpolation seems to be
reasonable.

5.4.3 Production and injection

To study the geophysical driving mechanisms of the deformation and to interpret

the interferometric data, production and injection data are provided by CFE. Pro-
duction data are given for the total flowrate (steam and water) as an average value
per day per well in tons per hour. The injection data are given in average values
per day per injection well in tons per second. Since 1985, the production has been
nearly constant. The production figures between April 1994 and March 1998 show
a constant total fluid mixture production of ~12 kT/h. Injection of brine started
in 1989, using injection wells located in the north of the evaporation pond. Injec-
tion is equally divided over all wells, and is nearly constant as well, approximately
2.3 kT/h.
Based on this information, and the knowledge of compaction-type subsidence (hence,
the surface subsidence is probably not seismic but largely continuous) it can be ex-
pected that the interferometric combinations reflect subsidence rates linearly pro-
portional to the temporal baselines.
 5.4 Cerro Prieto geothermal field 179

5.4.4 Interferometric results

Figure 2.12 (p. 44) shows all available SAR acquisitions over the Cerro Prieto
Geothermal Field from April 1995 to August 1997. To reduce the effect of geometric
decorrelation in the interferograms, ten acquisitions within a baseline band less than
200 m are selected. Two ERS tandem (1 day interval) acquisitions are available,
expected to show no deformation signal. The tandem pair with the longest base-
line is used to obtain an interferogram with significant topographic
information, while introducing a minimal amount of atmospheric signal in the dif-
ferential interferograms. Three-pass differential interferometry is used, scaling the
topographic interferogram and subtracting it from every deformation pair. The data
are filtered using the Goldstein and Werner (1998) filter to increase signal-to-noise
ratio.
Regarding the subsidence at the Cerro Prieto geothermal area, we expect that mainly
the vertical displacement is significant. Using eq. (5.1.1), one fringe of slant-range
deformation will correspond with approximately 30 mm vertical deformation.
Eight deformation interferograms are processed differentially, using the acquisitions
180 Chapter 5: Data analysis and interpretation for deformation monitoring

indicated by the solid dots in fig. 2.12. All interferograms are processed onto the
same reference grid, to enable different combinations and stacking analyses.
Figures 5.13 and 5.14 show all differential interferograms, using the tandem pair as
reference. The color bar shows the interferometric phase in radians, and the time
difference between the two acquisitions is indicated by the green bar. The contours
of the evaporation pond and the Cerro Prieto volcano, as well as the railroad and
major roads are superposed on the images for orientation.
Figure 5.13 shows the four 70 day interval differential interferograms. Especially
fig. 5.13A is a clear indication of the subsidence pattern, showing 0.5 cycles of de-
formation, corresponding with 1.5 cm subsidence, or 7.8 cm/year. The smoothness
of the interferometric phase suggests minimal atmospheric influence for this specific
pair. A relatively strong subsidence gradient is observed to the south of the evapo-
 5.4 Cerro Prieto geothermal field 181

ration pond, where a gradient of 1.2 cm over 1.5 km occurs. Temporal decorrelation
is clearly limited to some agricultural fields, whereas the geothermal area remains
relatively coherent. The subsidence pattern seems limited to the upper left part
of the image, in contrast with leveling data. Unfortunately, temporal decorrelation
outside the selected interferogram area is too large, limiting the interpretable area
to this crop. Figures 5.13B–D show three consecutive interferograms, each sharing
one SAR image with the next one. These three interferograms appear more noisy
than interferogram A. The small scale, relatively smooth variability in the inter-
ferometric phase is due to atmospheric signal. In first approximation this can be
concluded from the phase variation over the flat plain to the west of the evaporation
pond. Assuming homogeneous atmospheric conditions, such phase variations can
be expected over the subsiding areas as well. In section 5.4.9, the analysis of the
atmospheric signal is considered in more detail. Although the atmospheric signal
leads to phase variations of almost one cycle, the effects are limited in spatial scale,
compared to the subsidence pattern expected based on interferogram A. Therefore,
the subsidence pattern is still detectable by visual comparison, although an accu-
rate comparison of the magnitudes is more difficult. In the area containing the well
groups very localized “dots” with a different phase can be observed. These peculiar
effects may be explained by either very localized subsidence of the well installations,
due to their direct connection with the reservoir, or by atmospheric delay difference
induced by the very hot and humid steam escaping over the wells.
Figure 5.14 shows four interferograms with longer temporal baselines, 176, 245, 351,
and 491 days, respectively. The influence of increasing temporal decorrelation is
clear, diminishing the coverage of reasonable coherent data. Nevertheless, although
the data consist of isolated patches, it is still possible to follow the main fringes
visually. The contours of the fringes show the ellipse-shaped subsidence bowl. The
interferograms in fig. 5.14A and B are both combinations with the oldest image (20
May 1995). Since this common image appears as master in fig. 5.14A and as slave
in B, the fringe direction is reversed. It is clear to see that both interferograms do
not show the localized “sub-bowl,” which was observed in the 1994-1997 leveling,
cf. fig. 5.12. This suggests that this effect occurred between the 1994 leveling and
May 1995, or was an erroneous interpretation of the leveling, perhaps due to bench-
mark instability. Assuming the deformation rate estimated from fig. 5.13A to be
correct, the amount of phase cycles between the center of the subsidence ellipse and
the relatively stable area west of the evaporation pond can be extrapolated to these
intervals. This results in 1.3, 2.0, 2.6, and 3.5 fringes for fig. 5.14A, B, C, and D,
respectively. Especially for interferogram D, this implies that we would “miss” one
fringe by visual inspection only. Therefore, phase and patch unwrapping algorithms
need to be fine-tuned for this particular situation. Nevertheless, the results of these
four interferograms is consistent with those obtained from the 70 day interval pairs.

5.4.5 Stacking experiments

Stacking (adding) interferograms is an integer linear combination, as discussed in

section 3.5.3, see also Sandwell and Price (1998); Sandwell and Sichoix (2000). For
such combinations, performed using a complex multiplication to prevent wrapping
182 Chapter 5: Data analysis and interpretation for deformation monitoring

errors, phase unwrapping is not required. The result of adding deformation interfer-
ograms with common images is equivalent to calculating an interferogram with the
first and the last image. Here we show two examples of stacking.
Stack 1 (fig. 5.15A) is formed by the linear integer combination of three interfero-
grams 20/05/95-12/11/95, 12/11/95-21/01/96, and tandem interferogram 20/01/96-
21/01/96. This stack can be compared with one single interferogram 20/05/95-
20/01/96. The stack is shown in fig. 5.15A, the single interferogram is shown in
fig. 5.14B. The difference between both is not drawn since it is obvious that it shows
0 rad over the coherent areas, apart from some small differences introduced by the
filtering of the contributing interferograms. It is important to note that this example
does not imply that it is not necessary to calculate the interferograms over smaller
time intervals. Using this simple form of stacking, a decorrelated area in one of
 5.4 Cerro Prieto geothermal field 183

the contributing interferograms will contaminate the final result in the stack. An
improved way would be to use all partial results as well in the interpretation.
Stack 2 (fig. 5.15B) is formed by the linear combination of all interferograms spanning
the epoch between 20 May 1995 and 3 August 1997 (803 days). Obviously, such a
long temporal baseline results in a loss of coherence over the largest part of the
area. Nevertheless, there are some parts where a large phase gradient is still visible.
The expected total amount of fringes due to the subsidence is 5.7. Note that the
atmospheric influence in a stack is limited to the atmosphere in the first and the
last contributing SAR acquisitions, since intermediate atmospheric contributions
cancel in the differencing. Therefore, if possible, downscaling a stack covering a long
interval reduces the influence of these atmospheric contributions, whereas upscaling
of a short time interval exaggerates the atmospheric contribution.

5.4.6 Phase unwrapping strategies

Temporal decorrelation characteristics differ for various types of land use. This often
results in relatively coherent “islands” or patches in a decorrelated surrounding, see,
e.g., fig. 5.15A and B. As isolated patches are not connected, phase unwrapping
can be performed in every patch, but individual patches can have offsets. Two
(laborious) strategies for resolving phase ambiguities are tested on these data:

Johnny Appleseed 1 phase unwrapping refers to the identification and separate

unwrapping of all coherent patches, placing “seeds” for starting the phase un-
wrapping in a dense grid over the interferogram, followed by patch unwrapping
using a priori information.
1
Johnny Appleseed (John Chapman), was a legendary figure, immortalized in American liter-
ature by Vachel Lindsay (1879-1931), wandering from coast to coast through the U.S., planting
seeds of trees in every town. Also famous character in the Disney film Melody Times (1948).
184 Chapter 5: Data analysis and interpretation for deformation monitoring

Tie-point unwrapping. This approach is using local phase observations at

manually chosen positions, unwrapping them using a priori information or
first-order assumptions.

Johnny Appleseed strategy

Using the tree algorithm of Goldstein et al. (1988), branch cuts are defined connect-
ing residues. Phase unwrapping then starts at a (specified or random) seed-position,
and progresses within the branch cut limits. The Johnny Appleseed method places
a fine grid of seeds over the interferogram, and starts unwrapping for every seed that
is not already in an unwrapped patch. Hereby, unwrapped patches are numbered
consecutively until every seed position has been processed. This approach results in
an unwrapped phase for every individual patch. To unwrap the patches relative to
each other, each labeled patch can be offset by until a priori defined criteria are
satisfied. Often, manual editing of the patches is necessary, such as dividing, joining,
or eliminating operations. Figure 5.16 shows the initial results of Johnny Appleseed
unwrapping of the 20/01/96–05/01/97 interferogram. Figure 5.16A shows all identi-
fied patches, while fig. 5.16B represents the unwrapped results. In this specific case,
many patches are identified, making the procedure very laborious. Often, however,
interferograms consist of a few large patches, e.g. separated by waterways. In this
case, the reference to the water level makes the procedure fast and efficient, albeit
not automatic.
Tie-point unwrapping
The tie-point method uses manually selected tie-points in the wrapped interfero-
gram. The points are chosen based on coherence considerations, or assumptions
about the phase stability of individual or groups of resolution cells. The positions
and phase values are stored, and a two-dimensional interpolation is applied using a
 5.4 Cerro Prieto geothermal field 185

bi-cubic algebraic polynomial. Two examples of tie-point unwrapping are shown in

fig. 5.17. Knowledge about the deformation parameters (such as maximum phase
gradient per pixel per time epoch) is used to offset tie-point values by until
the total pattern satisfies deformation criteria. Since the source of fluid extraction
is situated between 1500–4000 m, the deformation signal will have relatively long
spatial wavelengths. Maximum deformation gradients are estimated from the level-
ing results in the order of 0.1 rad/pixel/yr, for a posting of 20 m. By subtracting
the re-wrapped interpolated result from the original interferogram, it is possible to
analyze the difference image. Local phase ramps indicate erroneous offsets, so
adjustments to these offsets can be applied.

Integer linear combinations

Since many of the analyzed interferograms show areas with severe temporal decorre-
lation, an iterative approach for the data analysis can be useful. Additional initializa-
tion assumptions can aid post-processing steps such as patch-unwrapping. Assuming
constant deformation in time, enables the use of relatively coherent (70 days) interfer-
ograms to extrapolate in time and predict deformation rates over longer time spans.
The use of integer multiplication factors makes phase unwrapping of the interfero-
gram to be scaled unnecessary. Drawback is the amplification of short wavelength
and atmospheric noise, when the scaling is larger than, say, a factor of three. This
approach can be used to offset isolated coherent patches by to get true absolute
phase values.
186 Chapter 5: Data analysis and interpretation for deformation monitoring

5.4.7 Forward modeling

To investigate the geophysical characteristics of the observed subsidence pattern

forward models are used (Hanssen et al., 1998a). A reasonable correspondence be-
tween the modeled and observed deformation pattern can provide coarse estimates
on the driving mechanisms of the subsidence. The subsidence is caused by vol-
ume contraction within the reservoir, which can be a result of a poroelastic and
thermoelastic mechanism. Point centers of contraction, or Mogi sources, are used
assuming that the earth can be modeled as an elastic half space (Mogi, 1958). Using
a single Mogi-source model a reasonable first-order fit of the one-year subsidence
rates is obtained, see fig. 5.19. The tie-point interpolated analysis of the 20/01/96–
05/01/97 interferogram (fig. 5.19A) is compared to the slant-range displacement as
inferred from a Mogi-source model at a depth of 5.9 km and a volume change rate of
(fig. 5.19B). The difference between the model and the obser-
vations (fig. 5.19C) shows residues of maximally 10% in the subsidence region. The
cross-section indicated by the blue line is shown in fig. 5.19D. We anticipate that a
better fit to the data could be obtained using a multiple Mogi-source model with a
shallower source depth and small volume change rate of, say,
as performed by Carnec and Fabriol (1999). On the other hand, a perfect fit using
a large number of Mogi sources may overestimate the accuracy and feasibility of the
model, especially regarding the influence of the assumptions on which this type of
modeling is based and its under-determined nature.

5.4.8 Deformation mechanisms

Fluid extraction from the pores of a volume of rock within the earth decreases the
pore pressure, which results in a contraction of the rock (Segall, 1985). As long as this
process is reversible, the mechanism is poroelastic. On the other hand, a decrease in
the volume can also be caused by the cooling of the reservoir. Although the removal
 5.4 Cerro Prieto geothermal field 187

of water from the reservoir does not produce a temperature drop, cooling of the
reservoir can be caused by the flashing of the water to steam due to its extraction
and the associated fast pressure drop (Mossop et al., 1997). As such, a thermoelastic
mechanism could also contribute to the observed deformation.

The reservoir dimensions of spatially and a 2 km thickness give an upper

bound for the volume of the reservoir of Hence the volume strain
rate, is about The volumetric coefficient of thermal
expansion, relates an increase in temperature, to volume strain, as
The volumetric coefficient of thermal expansion for greywacke at a similar
geothermal field (The Geysers, California) is assumed to be similar to the greywacke
here. Using this value, the strain requires a temperature decrease
of more than 1 K/yr to be thermoelastic. Compared with studies at The Geysers,
by Mossop et al. (1997), this value is far too large. At The Geysers (with very
similar production and injection rates, and 40% of the production,
respectively) a temperature decrease of 0.25 K/yr was inferred (Mossop et al., 1997).
Since the Geysers geothermal field is more efficient (higher steam production) and
cooler than Cerro Prieto (higher temperature decrease due to injection), the Geysers
temperature decrease can be taken as an upper bound for the expected temperature
decrease at Cerro Prieto.

This indicates that the observed deformation cannot be explained by thermoelastic

volume changes only. At most 25% of the deformation is thermoelastic, so the larger
part will be of poroelastic nature.

5.4.9 Identification of atmospheric signal using linear combinations

The identification of significant localized atmospheric signal is possible using lin-

ear combinations, or pair-wise logic (Massonnet and Feigl, 1995a). Figure 5.20
is a demonstration of this concept. Two interferograms (05/01/97–16/03/97 and
16/03/97–26/05/97, fig. 5.20A and B) share a common image acquired at 16/03/97.
Atmospheric signal during that acquisition will contaminate the subsidence signal
in both interferograms. Such anomalies are visible, e.g., in the center of the in-
terferograms, west of the evaporation pond, indicated by the arrows. Since the
16/03/97 acquisition is used as “slave” in the first interferogram and as “master”
in the second interferogram, the sign of the anomalies is reversed in both interfer-
ograms. Therefore, the sum of both interferograms results in complete removal of
these atmospheric anomalies (fig. 5.20C), whereas subtraction of the interferograms
results in an amplification of the atmospheric signal (fig. 5.20D) with a factor 2.
For this particular situation both interferograms cover 70 days, and the deformation
rate can be considered approximately constant in time according to the production
and injection data. Therefore, addition of the interferograms results in a doubled
(140 day) deformation signal, whereas subtraction removes the deformation signal
entirely.
188 Chapter 5: Data analysis and interpretation for deformation monitoring

5.4.10 Conclusions

Land subsidence over the Cerro Prieto geothermal area can be clearly detected by us-
ing radar interferometry. Systematic monitoring of the area using radar acquisitions
acquired every 35–70 days can provide a significant improvement of the knowledge
on the spatial distribution of subsidence as well as the subsidence rate and its tem-
poral fluctuations. Especially the monitoring of strong localized gradients in the
subsidence rates, see fig. 5.18, can be important for engineering purposes, e.g., to
avoid damage to infrastructural works. Furthermore, the combination of the radar
interferometric data with other geophysical parameters may lead to improved under-
standing of the thermoelastic or poroelastic characteristics of the subsurface. Based
on the interferograrns analyzed, the comparison with the leveling data shows that
the subsidence rate of the main bowl is approximately equal. However, a localized
 5.4 Cerro Prieto geothermal field 189

sub-bowl, identified in the 1994-1997 leveling (see fig. 5.12), is not observed in the
interferometric data. This implies that it is either a phenomenon which occurred
between the 1994 leveling and the oldest SAR image (May 1995), or an artifact in
the leveling, perhaps due to benchmark instability.
Problems in the interferogram analysis mainly occur due to temporal decorrelation
and atmospheric signal in the data. The temporal decorrelation leads to isolated
patches of coherent phase information, surrounded by decorrelated areas. Phase
unwrapping and patch unwrapping may yield ambiguous results. Atmospheric signal
may deteriorate single interferograms, making a sound interpretation of a single
interferogram cumbersome. The use of a priori information, as well as an increase
in the number of available images can help reducing these problems.
190 Chapter 5: Data analysis and interpretation for deformation monitoring

5.5 Izmit earthquake

On 17 August 1999, 3 AM local time, the area around Izmit, Turkey suffered a
strong earthquake: 7.5 on the Richter scale (Reilinger et al., 2000). The earthquake
was caused by an accumulation of stress during many decades around the North
Anatolian Fault (NAF) zone, where the Anatolian plate and the Eurasian plate
meet. It was the biggest earthquake since the 1967 Mudurnu Valley earthquake
of Ms=7.1, and had catastrophic consequences—Killing more than 17.000 people,
leaving thousands homeless, and causing $10–$25 billion in damage (Parsons et al.,
2000). Since the earthquake sequence of August–November 1999 can be regarded
as part of a largely westward propagation of ruptures along the NAF since 1939
(King et al., 1994; Stein et al., 1997; Stein, 1999), it is likely that it has increased
the stress on faults beneath Marmara Sea. This, in combination with an earthquake
history that indicates a seismic gap beneath Marmara Sea and tectonic loading
poses an increased earthquake risk for the city of Istanbul (Hubert-Ferrari et al.,
2000; Parsons et al., 2000), just a few hundred kilometers northwest of the Izmit
region.
This section contains a concise presentation of the coseismic interferometric data
(acquired before and after the earthquake), using the optimal interferometric com-
bination. It is not intended to give an in-depth geophysical analysis, but merely as
an example of an interferometric deformation signal that serves as a comparison for
the error sources discussed in the previous chapters.

5.5.1 Background

The North-Anatolian Fault forms the boundary between the Eurasian and the Ana-
tolian plates. Right-lateral strike-slip occurs along a nearly vertical fault plane.
Plate motion amounts up to 2.5 cm/yr (Hubert-Ferrari et al., 2000). The North-
Anatolian Fault branches of in a northern and a southern branch. The 17 August
1999 Izmit earthquake occurred on the northern branch, and ruptured a 110 km
long section.

5.5.2 Interferometric data

Izmit earthquake stimulated ESA to perform dedicated SAR observations over the
region, optimized for obtaining coseismic interferograms. ERS-1, which was in hi-
bernation mode at the time, had coincidentally been used one month before the
earthquake, and was switched on again the first possibility after the earthquake. An
orbit maintenance manoeuvre for ERS-2 was scheduled ahead, to enable near-zero
baseline conditions for the coseismic ERS-2/ERS-2 interferogram used in this study.
The perpendicular baseline was ~18 m, decreasing to nearly zero from near-range to
far-range, which corresponds with a height ambiguity of ~566 m. Especially at the
north of the Gulf of Izmit, topographic height variation is less than 600 m (Field-
ing et al., 1999), which results in less than one-fringe topographic contribution. To
the south of the Gulf, topography is higher, up to 1000 m, resulting in maximally
2 fringes topography.
5.5 Izmit earthquake 191
192 Chapter 5: Data analysis and interpretation for deformation monitoring

Figure 5.21A is the radar intensity image of ERS-2, acquired from 789 km height.
Brightness variation in the image shows topography and urban areas. The dotted
horizontal line indicates the approximate location of the North-Anatolian Fault.
Figure 5.21B is the coseismic (ERS-2/ERS-2) interferogram between 13 August and
17 September 1999. Since the topographic contribution is less than 2 fringes, the
interferogram shows primarily slant-range displacements due to the earthquake and
possibly postseismic deformation between 19 August and 17 September 1999.

5.5.3 Interpretation

Every fringe in fig. 5.21B corresponds with 28 mm of deformation, directed to the

radar or away from it. In this case a yellow-blue-red (YBR) color cycle means
relative deformation away from the radar, while yellow-red-blue (YRB) suggests
relative deformation towards the radar.
Approximately 27 fringes span the range between the Black Sea coast and the north-
ern shores of the Gulf of Izmit, corresponding with cm in slant range.
Traveling this range from North to South, the order of the colors is YBR, which
implies that the northern shores of the Gulf of Izmit have moved 75 cm away from
the radar, relative to a point at the Black Sea coast.
Here, the measured ~75 cm is the deformation in the radar line-of-sight. Since the
dominant part of the strike-slip deformation along this fault system is horizontal
(Barka, 1999), we find with eq. (5.1.1) a horizontal component in the azimuth look
(ALD) direction of ~7.2 cm/fringe. The decomposition of this horizontal displace-
ment vector in a north and east component is not possible without any a priori
information. Nevertheless, since the fault is oriented nearly perpendicular to the
satellite track, the ALD component will be of the same order of magnitude as the
absolute horizontal deformation vector. As a result, the northern shore of the Gulf
of Izmit has moved ~1.95 m in ALD direction, relative to a reference point at the
Black Sea coast.
From the interferogram it can be concluded that it is very difficult to count the
fringes near the fault. This is caused by two mechanisms. First, near to the fault
the deformation rate is probably to high, so that undersampling occurs. This is
effectively the upper deformation gradient of see section 2.4.3 and fig. 2.11.
Second, temporal decorrelation in the rupture zone area is extreme, resulting in a
loss of coherence over many of the damaged area. Therefore, extrapolation of the
fringes, or use of additional observations is necessary to obtain deformation rates
near the fault. From field observations it is known that horizontal displacement
between 2 and 3 m has occurred near the fault. Near the Gulf of Izmit observed
deformations amounted up to 4–4.5 m.

5.5.4 Analysis of the observed pattern

In fig. 5.22 a first order, coarse interpretation of the fringe pattern is presented. The
fault plane represents the boundary between the Eurasian plate in the north and the
Anatolian plate at the south side. The dotted line indicates the shortest distance
 5.5 Izmit earthquake 193

between the two squares. The general plate movement causes the two squares to
move away from each other, parallel to the fault plane (fig. 5.22A and B). Without
direct slip or creep along the fault, the dotted line will bent significantly with time.
Just before the earthquake, the solid line indicates the shortest distance to the
fault plane. Immediately after the earthquake, when both plates have slipped, the
situation is as in fig. 5.22C. As the accumulated strain has been (partially) released,
the red line has moved to its new position. The blue line now indicates the amount
of deformation between shortly before the earthquake and shortly after. In fact, the
observed deformation pattern can be regarded as a mirrored version of the strain
pattern before the earthquake; cf. the red line in fig. 5.22B with the blue line in
fig. 5.22C.
As long as the fringes in the interferogram are parallel at both sides of the fault, the
two sides have moved parallel to each other. At those areas the strain which was
accumulated over the years is relieved at least partially. In the interferogram it can
be observed that the color patterns are not parallel the entire range of the fault: in
194 Chapter 5: Data analysis and interpretation for deformation monitoring

the western part they bend towards the Gulf of Izmit, both on the northern and the
southern side of it. Assuming that this behavior continues through the crust under
the Gulf of Izmit, this implies that both sides of the fault have not moved with
respect to each other. Since the fringes cannot be followed across the water, there is
can be an inaccuracy of several fringes in this assumption, but still the deformation
is significantly less than closer to the main epicenter. Since the end of the rupture
zone was under water, it could not be clearly detected. The direction of the fringes,
however, indicates that the end of the rupture zone was near the left side of the
interferogram Assuming that at this position the rupture zone ended, a large part
of the strain will still be apparent, which could be an indication for an increased
probability for future earthquakes, which is in agreement with general understanding
of stress induced triggering (King et al., 1994; Stein et al., 1997; Stein, 1999). Proper
ways to localize areas where strain accumulation is still in progress are essential in
hazard monitoring and risk assessment.

5.5.5 Conclusions

This brief presentation of the coseismic interferogram of the Izmit earthquake shows
some of the pros and cons of the technique. The advantages are clearly the increased
spatial coverage with respect to conventional geodetic techniques, see Reilinger et al.
(2000) for an overview of these data for the Izmit earthquake. Furthermore, the pos-
sibility to combine an archived SAR acquisition prior to the earthquake with a new
acquisition after the event (including orbit optimization) results in a reduced demand
on monumentation and in situ survey networks. Nevertheless, the InSAR data only
yield line-of-sight deformations, which demand either reliable assumptions on the
physics of the deformation, or a complementary combination with other geodetic
or geophysical techniques. Temporal decorrelation can eliminate the possibility of
a coherent interferometric combination. For example, if the earthquake would have
happened a few months later, in the snow season, the chances to obtain a coherent
interferogram would have been reduced dramatically.
It needs to be emphasized that the study of the geophysical interpretation of the
Izmit SAR data is still in progress, aiming at the development of a best-fit model to
describe this earthquake, see Riva et al. (2000).

5.6 Conclusions

Before commenting on the general feasibility of radar interferometry for deformation

studies, it is necessary to make a, perhaps more philosophical, comment. In the case
studies presented here it was already known that a deformation occurred, and the
problem was approached from a survey-strategy point of view. It needs to be stressed
that the technique has an unsurpassed quality to serve as a monitoring tool, even
if the quantitative results are not uniquely interpretable. In contemporary geodesy,
the goal of obtaining more accurate measurements is slowly shifting—in terms of
user requirements—towards simply more spatially distributed observations, acquired
more often. In many cases, areas which suffer from deformation (earthquakes, sub-
sidence, landslides, volcano dynamics) are not monitored at all, or observations are
 5.6 Conclusions 195

only available after a catastrophic event. A flexible method for acquiring new radar
images combined with an archive of historical data may be the only means to analyze
this type of situations, see for example volcano studies in Alaska (Lu et al., 1997) or
at the Galapagos (Amelung et al., 2000a,b).
To evaluate the feasibility of radar interferometry for a specific application, especially
regarding the influence of atmospheric signal in the data, it is necessary to compare
the scale characteristics of the deformation phenomena with those of the atmospheric
signal. Figure 5.23 shows a sketch of the expected scales for the examples discussed
in this chapter. It can be regarded as complementary to the instrumental evaluation
shown in fig. 2.11, cf. p. 41. In fig. 5.23, spatial scale is expressed on the horizontal
axis, whereas the vertical axis shows (i) the estimated magnitude of the slant-range
deformation signal in range, and (ii) the square-root of the atmospheric structure
function, or the standard deviation of the atmospheric signal between two points
with a certain spatial separation. The dashed line represents an approximation
of the square-root of the structure function based on the experimental results in
chapter 4. The solid lines in the four plots represent the characteristic scales of the
discussed deformation processes. For Cerro Prieto, the main subsidence bowl covers
196 Chapter 5: Data analysis and interpretation for deformation monitoring

scales ranging from 1–15 km. The magnitude depends on the time interval between
the SAR acquisitions, as this is a nearly linear phenomenon. From the analyzed
data in section 5.4 it is clear that the subsidence signal is detectable in a 70 day
interferogram, although atmospheric anomalies may contaminate scales less than a
few kilometers.
For the subsidence problems in Groningen, the spatial scales are larger, say 5–30 km.
Since the subsidence rate is slow, less than 1 cm/yr, it takes a considerable amount
of time before the subsidence signal dominates over the atmospheric noise level. Ob-
viously, this time period strongly influences the amount of temporal decorrelation.
The earthquake in Turkey shows deformation signal over a wide range of scales, up
to ~1.5 m can be observed in slant range. Over smaller scales, the amount of defor-
mation decreases, apart from the rupture zone, where very strong signal is apparent
over just a few meters. Note that the latter is not measurable by interferometry
due to the steep phase gradient, see fig. 2.11. Over all scales, the deformation signal
is stronger than the atmospheric signal. Although this implicates that the defor-
mation signal is detectable, it needs to be stressed that the atmospheric signal still
contaminates the data. Finally, the corner reflector experiment (~1 cm deformation)
only concerns a localized group of a few pixels, indicating a negligible atmospheric
contribution.
Simple analyses as sketched above may serve as an important tool to guide the
decision whether to apply spaceborne repeat-pass radar interferometry for a specific
deformation problem.
 Chapter 6
Atmospheric monitoring

This chapter presents first results of Interferometric Radar Meteorology (IRM),

showing how radar interferometric delay measurements can be used to infer high
resolution maps of integrated atmospheric water vapor. Maps of the water va-
por distribution can be readily related to meteorological phenomena, and suggest
that radar observations can be used to study atmospheric dynamics and may
contribute to forecasting. The theory and sensitivity analysis of the various
atmospheric refractivity contributions is discussed. IRM results are validated
using conventional meteorological techniques and GPS observations.
key words: Atmosphere, Troposphere, Ionosphere, Water vapor, Meteorology

The spatial distribution of water vapor in the earth’s atmosphere is important

for climate studies, mesoscale meteorology, and numerical forecasting (Rind et al.,
1991; Emanuel et al., 1995: Crook, 1996). Local concentrations of water vapor
vary between 0–4%, which makes it one of the most varying atmospheric con-
stituents (Huschke, 1959). Although extensive ground-based and upper-air sound-
ing networks are routinely used, these measure the water vapor distribution only at
coarse scales. Even spaceborne radiometer observations, such as Meteosat or GOES,
have spatial resolutions that are currently too coarse for, e.g., quantitative precip-
itation forecasting or cloud formation studies. Moreover, radiometric water vapor
observations from these satellites usually originate from atmospheric layers above
3 km, due to the strong absorption by water vapor (Weldon and Holmes, 1991).
This often restricts quantitative interpretation to upper-tropospheric moisture dis-
tribution (Schmetz et al., 1995).
These limitations form the main source of error in short-term (0–24 hr) precipitation
forecasts (Emanuel et al., 1995). For example, thunderstorm initiation and strength
are sensitive to spatial and temporal variations in moisture of the order of 1 g/kg
(1.2 hPa) and temperature of 1–3°C (Crook, 1996; Mueller et al., 1993). Such
variations are common on a 1-km spatial scale (Weckwerth et al., 1997). SAR,
however, provides horizontal resolution as fine as 10–20 m over a swath typically
100 km wide.
It is widely accepted that spaceborne geodetic observations such as GPS and VLBI
provide an independent source of meteorological information by analyzing the prop-
agation delay of the electromagnetic signals. Availability of these data on a routine
198 Chapter 6: Atmospheric monitoring

basis is improving numerical forecasting and mesoscale meteorology. In this chapter,

we will show that radar interferometry can be used for meteorological purposes as
well, even though the configuration is currently not suited for operational meteo-
rology. Nevertheless, the combination of high accuracy and fine spatial resolution
with the huge archive of interferometric datasets can be readily used to provided
unsurpassed insights in atmospheric dynamics.
The atmosphere below about eighty kilometers naturally divides into two layers, the
troposphere and the middle atmosphere, where the latter consists of the stratosphere
and the mesosphere (NorthWest Research Associates, 2000). The troposphere is
destabilized by radiative heating and by vertical wind shear near the surface. While
radiative heating leads to convection, the vertical wind shear leads to baroclinic
instability that manifests itself in synoptic-scale storms and frontogenesis. Unlike
the troposphere, the overlying middle atmosphere is relatively weather-free. More
precisely, the meteorological phenomena of the middle atmosphere do not resemble
anything commonly experienced in the troposphere. The stratosphere is surpris-
ingly dry, so cloud formation is almost nonexistent apart from noctilucent and polar
stratospheric clouds (NorthWest Research Associates, 2000). Therefore, signal delay
in the neutral part of the atmosphere is mainly originating from the troposphere.
Mesoscale Shallow Convection (MSC) represents a specific class of atmospheric sys-
tems that could benefit from interferometric radar observations. Mesoscale systems
have horizontal length scales between ten and a few hundred km. Shallow systems
occupy the lower 1 to 2 km. This layer is referred to as the planetary boundary
layer; the lowest part of the atmosphere, where the earth’s surface has a profound
effect on the properties of the overlying air. In this framework, convection repre-
sents mass motions of air resulting in vertical transport and mixing of its properties,
which is a principal means of transporting heat. Atkinson and Zhang (1996) state
three reasons why MSC are important. First, the transport of moisture, heat, and
momentum to the free atmosphere are crucial for general circulation. Second, the
net radiation budget of the earth, important for understanding the earth’s climate
system, is significantly influenced by stratocumulus clouds. MSC is related to the
occurrence and forming of such clouds. Finally, as long as the process of MSC is
not fully understood, there is a gap in the understanding of the dynamics of the
planetary boundary layer.
This chapter is intended to be self-contained, allowing for a direct reference by, e.g.,
meteorologists and for atmospheric studies. It starts with a brief description of
the theory necessary for the meteorological interpretation of radar interferometric
data. It will be argued that for short spatial scales, say less than ~50 km, the
dominant atmospheric signal is due to the water vapor distribution. In section 6.2,
the influence of water vapor, temperature, and liquid water (droplets) is discussed.
Section 6.3 elaborates on the effects of the ionospheric electron density. In a previous
study (Hanssen, 1998), 26 interferograms have been analyzed and interpreted from
a meteorological point of view. In sections 6.4–6.7, a number of these case studies
are discussed in more depth, focusing on the interpretation and validation of the
observations. Section 6.8 concludes this chapter summarizing the potential and
limitations of water vapor mapping with radar interferometry.
 6.1 Theory 199

6.1 Theory

A SAR image contains information on the path length between the radar antenna
and the resolution cells on earth, see chapter 3 for a more elaborate discussion. The
interferometric combination of two radar images with a temporal separation of 1 day
provides a sensitive tool to measure these path length differences as a fraction of the
radar wavelength, which is 5.66 cm for C-band radar used here. Conventionally,
path length differences can be attributed either to topographic height differences,
depending on the relative positions of the satellites (Zebker and Goldstein, 1986), or
to surface deformation, depending on the time interval between the two observations
(Gabriel et al., 1989; Massonnet et al., 1993, 1995). However, effective path length
variations are also caused by radar signal delay variability within the imaged area,
due to the heterogeneous refractivity distribution in the atmosphere. Atmosphere
induced distortion has been observed in radar interferograms (Goldstein, 1995; Mas-
sonnet and Feigl, 1995a; Tarayre and Massonnet, 1996; Hanssen and Feijt, 1996;
Zebker et al., 1997) but has typically been treated as noise.
Signal delay, in seconds, is equivalent to an excess path length by multiplication
with the speed of light in vacuum. The excess path length can be directly obtained
by integrating over the (dimensionless) refractivity along the line of sight. Over
small spatial scales, the variation in the integrated refractivity is mainly due to the
spatial variation of water vapor during the two image acquisitions. To a lesser degree
temperature, liquid water, and pressure gradients influence the delay variation.
Delay measurements observed by space-geodetic techniques can be used to derive
precipitable water vapor in the atmosphere (Saastamoinen, 1972; Hogg et al., 1981).
Precipitable water vapor is the amount of vertically integrated water vapor and can
be expressed in or as the height of an equivalent column of liquid water
in meters. GPS measurements provide temporal variations in precipitable water
vapor at one position, see e.g., Bevis et al. (1992). Using InSAR, the data reflect
spatial variations in precipitable water vapor during the two image acquisitions.
Figure 6.1 shows the geometric configuration of the radar acquisition and the GPS
zenith measurement.

6.1.1 Interferometric phase review

The SAR of the ERS-1 and ERS-2 satellites provides an amplitude and a phase
value for every resolution cell of approximately 4 × 20 m. Information on the path
length between the radar antenna and a ground resolution cell is contained in the
phase measurement. Unfortunately, the phase observation of resolution cell in
a SAR acquisition at is a superposition of a number of contributions:

where is related to the geometric distance and to the signal prop-

agation velocity variations. Most important, the scattering component, , is
the contribution of many arbitrary scatterers in the resolution cell, which add up to
produce a uniform probability density function (PDF) for As a consequence,
200 Chapter 6: Atmospheric monitoring

the sum of the PDF’s of all components in eq. (6.1.1) will have a uniform distribution
as well, and no useful phase information can be obtained.
In the repeat-pass interferometric combination, two SAR images, acquired at differ-
ent times, are accurately aligned and differenced, which yields the interferometric
phase:

consisting of a differenced geometric component propagation component

and scattering component If the second satellite orbit is sufficiently
close (< ~600 m) to the first and the physical scattering characteristics remain
constant the scattering component in each interferometric phase
observation will be eliminated in the differencing. In that case, useful information
may be retrieved from the resulting difference phase image or interferogram. The
assumption of a stationary scattering component is the limiting factor for the appli-
cation of SAR interferometry. For example, over water or rapidly changing surfaces
it is not possible to obtain coherent phase observations. Over many agricultural
areas, as in the test site in this study, phase noise increases with increasing time
intervals or due to anthropogenic activity. During the ERS-1 and ERS-2 “tandem
mission,” which lasted from August 1995 to April 1996, SAR images of the same
 6.1 Theory 201

area were acquired with a sampling interval of 24 hours. This short time interval
ensures a sufficiently high correlation between consecutive acquisitions over most
land surfaces.
In order to analyze the propagation component of the interferometric phase, the
influence of the geometric component needs to be eliminated. Geometric phase dif-
ferences are caused by either a change in satellite position or a coherent change in
the position of the scatterers on earth, between the two acquisitions. A difference
in satellite positions will measure topographic height variation in the SAR image.
Using a reference elevation model, a synthetic topographic interferogram can be
constructed that can be subtracted from the observed interferogram, resulting in
a so-called “differential,” topographic-free, interferogram (Massonnet et al., 1993).
Other variations in the geometric component, for example, due to surface deforma-
tion, can be safely ignored for these short time intervals. Therefore, observed phase
gradients in the differential interferogram can only be attributed to propagation
delay variability and residual trends due to the inaccuracy of the satellite posi-
tion during the acquisitions. Finally, the interferometric phase, which is originally
“wrapped” to the interval is unwrapped using dedicated phase-unwrapping
algorithms, see, e.g., Goldstein et al. (1988); Ghiglia and Pritt (1998).

6.1.2 Interferometric delay analysis

After obtaining the differential interferogram, the observed phase differences can be
interpreted as (i) the spatial delay variation between the radar antenna and millions
of pixels in the interferogram and (ii) the difference between the two generally uncor-
related states of the atmosphere during the two SAR acquisitions. Due to satellite
orbit errors and the wrapped nature of the phase observations, it is only possible
to measure the lateral variation of the delay, rather than the total delay. The delay
variation between pixel p and q is directly related to the interferometric
phase difference by

for C-band. Mapping the incident delay variation to zenith values can be achieved
by using a mapping function, e.g., a simple cosine function:

with incidence angle (look angle plus the earth’s curvature, see fig. 2.9) varying
between 19° and 23°, see fig. 6.1. Such a simple mapping function is sufficiently
accurate for steep incidence angles (Bean and Dutton, 1968).
As discussed in section 4.2, the standard deviation of the phase, is derived from
the coherence that is, the amount of correlation between the two SAR images,
with For we find Using eq. (6.1.3) and (6.1.4), and
averaging over 5 pixels to obtain 20 × 20 m resolution cells yields a formal accuracy
of the zenith delay (vertically integrated refractivity) observations of
202 Chapter 6: Atmospheric monitoring

Additional spatial averaging to a ground resolution of approximately 160 × 160 m

yields a phase standard deviation and consequently
Possible systematic errors in these delay observations are manifested as long-wave-
length gradients in the interferogram, caused by inaccuracies in the satellite’s po-
sition during image acquisition. These additional tilts are removed from the in-
terferogram before analyzing the delay differences. This procedure also removes
long-wavelength delay gradients caused by pressure and temperature gradients or
gradients in the ionospheric electron density.
The technique now reveals slant delay differences or integrated refractivity along
every path between the antenna position, and the resolution cells on earth. The
relation between zenith delay observed at resolution cells and and the refractivity
distribution during SAR acquisition can be written as

where is the dimensionless refractivity. The zenith delay variation

observed in the interferogram can be defined as

The refractivity N can be written using two equivalent expressions (Smith and
Weintraub, 1953; Kursinski, 1997; Hanssen, 1998). The first, original, expression
is

where is the partial pressure of dry air in is the partial pressure of wa-
ter vapor in and T is the absolute temperature in Kelvin. The constants
and have been first determined by Smith and Weintraub (1953), but
results from Thayer (1974) are also commonly used. We use
and These constants are considered
to be accurate to 0.5% of N (Resch, 1984). The electron density per cubic meter
is expressed by is the radar frequency (5.3 GHz), and W is the liquid water
content in The first term at the right-hand side of eq. (6.1.7) is often labeled
as the dry term, the following two terms (in brackets) as the wet terms. The last
two terms are the ionospheric term and the liquid term respectively.
Assuming that the total atmospheric pressure and using the equation of
state1 a second, equivalent, expression for N can be obtained (Davis et al., 1985):

1
Note that in this case, the equation of state should involve the virtual temperature see
Haltiner and Martin (1957).
 6.1 Theory 203

where P is the total atmospheric pressure in and

with and In eq. (6.1.8), the first

term at the right-hand side is referred to as the hydrostatic term, instead of the dry
term (Davis et al., 1985).
If we consider the contribution of the hydrostatic term in eq. (6.1.8) on the zenith
delay in eq. (6.1.5), we may write for point at time using the ideal gas law

Since the measured total surface pressure can be written as

where is the local gravity at the center of the atmospheric column (Saastamoinen,
1972), which is approximated by

dependent on the surface height and latitude the hydrostatic delay is obtained
from a simple barometric measurement

Using these parameters and the surface pressure measured with an accuracy
of 0.4 hPa or better, this delay can be predicted with an accuracy of 1 mm or
better (Bevis et al., 1996). For the test sites used in this study,
hence in the order of 2.3 m. For comparison, the wet delay is less
than ~0.3 m (Elgered, 1982).

6.1.3 Discussion

Standard observations of surface pressure within a typical interferogram area of

100 × 100 km show (i) minimal spatial variation, usually less than 1 hPa, and (ii)
a smooth behavior. This is due to the large scales of high and low pressure zones.
Therefore, the hydrostatic delay observed in interferograms is smooth as well and
manifests itself usually as a phase trend of maximally a few millimeters over the
entire interferogram. Such phase trends are likely to occur due to orbit errors as
well, see fig. 4.23, and are often corrected for using tie-points. As a result, the
hydrostatic delay has a very limited influence on interferograms, and can be safely
ignored for spatial scales of ~50 km and smaller. If spatially distributed surface
204 Chapter 6: Atmospheric monitoring

pressure measurements are available, these can be used to correct for the hydrostatic
delay component.
The wet part of the delay or refractivity, parameterized by temperature T and the
partial pressure of water vapor e, is much more spatially variable then the hydrostatic
delay, while the variations have a considerable effect on the observed phase delay.
The sensitivity of the observed delays to the wet part of the refractivity is discussed
in the following section.

6.2 Refractivity sensitivity analysis

This section elaborates on the interpretation of lateral delay variations, emphasizing

on the possibilities to distinguish the contribution of temperature and moisture
(partial water vapor pressure) to the total refractivity. To select feasible atmospheric
parameters, the partial pressure of water vapor is ranging from 0 hPa to 30 hPa,
and the temperature is ranging from to 40°C. We can express the relative
humidity, RH, by

where the saturation pressure for a certain temperature, can be found using the
Clausius-Clapeyron equation (Stull, 1995)

where T is the temperature in Kelvin, and L is the

latent heat. Over a flat water surface, which is the latent
heat of vaporization. In this case, cloud droplets are being formed. The relative
humidities for these ranges of temperature and partial water vapor pressure are
represented by the iso-lines in fig. 6.2.

6.2.1 Clear air refractivity decomposition

Under the assumption that we can ignore or derive the hydrostatic component of the
delay as discussed in the previous section, we can concentrate on the wet component.
Using the clear-air wet tropospheric part of eq. (6.1.8), the partial derivatives can
be derived to determine the sensitivities:

Note that a temperature increase yields a decrease in refractivity. Hence a localized

area with a higher temperature will show up as a relative increase in propagation
velocity, whereas a localized area with higher water vapor pressure will show up
as a relative decrease in propagation velocity. Figure 6.2 shows the ratio
 6.2 Refractivity sensitivity analysis 205

In other words, we compare (i) the sensitivity of for a 1 hPa

change in the partial pressure of water vapor e with (ii) the sensitivity of for
a 1°C change in temperature T. If this ratio |R| > 1 the refractivity is R times
more sensitive for a 1 hPa change in e than for a 1°C change in T. On the
the partial pressure of water vapor is plotted, on the the temperature. By
combining a specific e and T value it is possible to evaluate the sensitivity ratio.
Relative humidity isolines are superposed to show the likelihood of a specific e–T
combination. From fig. 6.2, we may conclude that the sensitivity ratio R has wide
range of values. Although some very high ratio’s are apparent for very dry air (less
than 5 hPa water vapor pressure), we can safely state that 4 < \R\ < 20. The ratio
is lowest for high temperatures combined with high partial water vapor pressures.
These conditions are likely in hot and humid regions, especially near the earth’s
surface. For cloud development, reaching higher altitudes, temperatures drop with
~6.5 K/km, which increases the ratio.
As a result, the wet part of the refractivity is in any scenario at least 4–20 times
as sensitive to a 1 hPa change in the partial pressure of water vapor than to a 1°C
change in temperature. Moisture variations of are common even
on a 1-km spatial scale (Weckwerth et al., 1997). At 0°C, such a variation already
produces 6 mm delay per vertical km.

6.2.2 Propagation through clouds and droplets

From the signal delay obtained from an interferogram, precipitable water vapor dif-
ferences can be obtained, under the assumption that the atmosphere is a clear gas.
However, the occurrence of clouds (liquid water droplets in a saturated environment)
and precipitation demonstrates that this assumption is obviously often not correct.
206 Chapter 6: Atmospheric monitoring

In many space geodetic techniques where the total wet delay is measured, the con-
tribution of the droplets is often neglected, since it is only a small part (estimated
1–5% by Kursinski et al. (1997)) of the total wet delay, hence 1–15 mm. InSAR,
however, is effectively blind for the total delay, but observes lateral differences in
delay instead. Although maximum lateral differences of 10–12 cm have been ob-
served, see Hanssen (1998), the sensitivity of the radar measurements allows for
detecting mm-scale fluctuations. Therefore, it might not be sufficient to restrict to
the clear gas hypothesis. In the following, we investigate the influence of droplets
and comment on the sensitivity of the precipitable water vapor estimate.
Phase delay induced by droplets
The phase shift of spaceborne C-band radar pulses caused by liquid water (droplets)
can cause limited additional signal delay in the interferometric observations. The
interaction of radar waves with droplets is a forward scattering problem: the wave
induces a dipole moment in the droplet, which will act as a secondary wave front.
After passing the droplet, the principal (undisturbed) wave front will interfere with
the secondary wave front and hereby cause a phase shift.
Hall et al. (1996) have listed the liquid water content, W, of clouds, see table 6.1.
The liquid water content is the particle number density times the volume per particle
times the density of liquid water. Its maximum is usually found at ~2 km above
the cloud base, and can be related to the dielectric refractivity using the Clausius-
Mossotti equation, (Solheim et al., 1997):

where is the permittivity of water and is the density of liquid water. This
relation is independent of the shape of the cloud droplet (Born and Wolf, 1980).
Although the permittivity of water is a weak function of temperature, it is possible
to approximate eq. (6.2.4) to within 1% by (Solheim et al., 1997):
 6.2 Refractivity sensitivity analysis 207

Since the value of reflects the difference with the vacuum refractive index,
eq. (6.1.8) can be used to determine the additional delay caused by the liquid
water in clouds:

where L is the thickness of the cloud layer in km. Note that cloud-droplet refrac-
tivity is in fact dispersive. However, since the dispersive part of the refractivity is
much smaller than the non-dispersive part it can be ignored for C-band frequencies.
Table 6.1 shows the zenith delay for four cloud groups. For repeat pass SAR inter-
ferometry, stratiform clouds and ice clouds do not cause large phase disturbances,
due to their large horizontal extent and small additional delay. However, especially
the cumulus type of clouds can result in a significant additional phase delay, as they
have a relatively limited horizontal size combined with a large vertical height and
liquid water content.

6.2.3 Comparison of water vapor and liquid water delay

Neglecting the hydrostatic and the ionospheric term in eq. (6.1.8) we find the refrac-
tivity expressed in the wet (vapor) and liquid terms:

If precipitable water vapor is defined as

where is the density of liquid water and is the density of wa-

ter vapor, we can relate the precipitable water vapor to the delay observed in the
interferogram, eq. (6.1.8), using the equation of state as
208 Chapter 6: Atmospheric monitoring

where and is the mean temperature of the column

containing the water vapor, see section 6.6. This can be written as

If the amount of precipitable liquid water is given by

this can be related to the delay induced by the droplets by

Using a typical value for (Bevis et al., 1996), we obtain the following
approximations

For a cumulus congestus, cf. table 6.1, Using a cloud depth of 4 km,
and using eq. (6.2.11), we find PLV = 4 mm. Following eq. (6.2.14), this yields
5.6 mm of zenith signal delay, or 0.2 phase cycles in the interferogram, which is
very well detectable. For other cloud types, such as stratiform clouds, ice clouds,
and small cumulus clouds, the effects are less than 1 mm, under usual atmospheric
circumstances. It is important to note that cloud droplets will only occur in a
saturated air mass, and therefore the water vapor concentration will be relatively
high. Hence, for the delays observed in the interferogram water vapor will be the
main driving force.

6.2.4 Discussion

These theoretical studies show that the sensitivity of the tropospheric refractivity is
highest for spatial variation in water vapor content, between 4 and 20 times greater
than for temperature variation. This sensitivity, in combination with the fact that
water vapor exhibits a strong spatial variability, supports the interpretation of small
scale (< 50 km) phase delay variations in terms of water vapor. Nevertheless, the
results show that temperature effects and liquid water in the troposphere cannot be
ignored, and result in deviations in the quantitative analysis of the delays. Regard-
ing liquid water, especially cumulus congestus clouds, cumulonimbus clouds, and
precipitation will modulate the observed water vapor quantities, even though the
delay differences due to the strong water vapor gradients will be dominant. Com-
puting precipitable water vapor without taking the liquid effects into account may
result in an overestimation of PWV of, say, less than 10%.
For the sake of completeness it needs to be stressed that, applying radar interfer-
ometry over a relatively flat area, there is no sensitivity for the vertical profiles of
e, P, T, and W as long as changes in these profiles are homogeneous for the image.
 6.3 Ionospheric influence 209

There is only sensitivity for lateral inhomogeneities in the refractivity N, hence,

lateral inhomogeneities of P, T, e, and W. Note that a different vertical layering
during the two SAR acquisitions will, however, influence the interferogram if sig-
nificant topography is present, see section 4.8. The signatures of these effects will
have strong correlation with the topography, which provides a possibility to identify
them. For the test sites analyzed here, with height variation within a range of 100 m,
no topographic induced effects are expected nor observed.

6.3 Ionospheric influence

The ionosphere can be physically characterized by the ability of external sources to

“knock-off” electrons from atoms, hereby creating free electrons that are not bound
to their remaining ions. These external sources are mainly the solar ultraviolet
radiation and energetic electrons of solar and magnetospheric origin. The number
of (free) electrons is represented by the electron density, in electrons per
Although the electron density is temporally and spatially variable, the ionosphere
is often treated as a spherical shell between 60 and 600 km height with a constant
electron density in height, homogeneously varying in time. Radio signals traversing
through the ionosphere are dispersively delayed along their paths by interactions
with the free electrons. The dispersive (frequency dependent) nature of the delay is
the key for e.g. dual-frequency GPS to estimate the ionospheric (or plasma) delay.
Physically, the ionospheric delay is a path integral through the ionospheric electron
density. This integral therefore has the dimension electrons per and is commonly
known as the Total Electron Content or TEC. If the path integral is taken in the
zenith direction it is known as the Vertical TEC (VTEC). To achieve comprehensible
quantities, one TEC unit (TECU) is equivalent to Using dual
frequency GPS observations collected by the IGS network (International GPS Service
for Geodynamics), maps of the Vertical TEC can be produced, see fig. 6.4, showing
two-hourly vertical TEC over Europe, provided by the DLR Remote Sensing Ground
Station Neustrelitz.
The delay of a radio wave, propagating through the ionosphere, is different for the
phase and the group velocity. Phase delay depends on electron content and affects
carrier signals. Group delay depends on dispersion in the ionosphere as well, and
affects signal modulation, i.e. the codes on GPS signals. The phase and group delays
are of the same magnitude but have an opposite sign. The group signal delay
through the ionosphere can be accurately represented by a series in the reciprocal
of the carrier frequency

Dual-frequency GPS receivers observe a differential delay between the two

frequencies L1 and L2, which enables estimation of the unknown A, B, and C
coefficients.
210 Chapter 6: Atmospheric monitoring

6.3.1 The mapping function

Radio signals that are not incident from zenith direction are generally delayed (or
advanced) as well as bended. To derive the mapping function, M, that accounts for
these effects, an idealized single-layer approximation is often applied. The altitude at
which a ray between satellite and the earth’s surface pierces the layer with maximal
electron density is defined as the effective ionospheric height The projection
of this point on the ground is labeled the sub-ionospheric point, indicated by the
subscript ’sp’. In the single-layer approximation, it is assumed that the whole elec-
tron density profile, which varies with altitude, is reduced to a very narrow spherical
layer at this effective ionospheric height. In general, an effective ionospheric height
is used.
For calculating the mapping function we need the zenith angle of the rays at the
 6.3 Ionospheric influence 211

sub-ionospheric point at height The mapping function is then defined as

where the cosine of the zenith angle can be approximated by

where is the earth’s radius and is the elevation angle relative to the horizon.
Using the ERS SAR incidence angle the earth radius of
6371 km and effective ionospheric height the value of the mapping
function is M(23°) = 1.07531.

6.3.2 Ionospheric delay

The zenith ionospheric range error for point p at is in first approximation

given by the formula:

where is the zenith range error in meters, is the frequency of the signal
and TEC is the Total Electron Content. The factor (Jakowski
et al., 1992). This is the ionospheric delay in the zenith direction. TEC values may
vary between at night to at the minimum of the solar cycle to
at the solar maximum. Note that the sign of the factor is negative,
i.e., an increase in TEC results in a phase advance.
For a SAR acquisition, under an incidence angle of approximately 23° and a radar
frequency of 5.3 GHz, this yields

Taking the derivative of this function and using TEC-units (TECU), we obtain for
C-band SAR

In other words, for a repeat-pass interferometric cycle of 28 mm, a change in 1 TECU

over the scene yields a phase ramp of approximately half a cycle. Of course, we have
to take into account that an interferogram is the difference of two SAR acquisitions
during different days, but at the same local time. The status of the ionosphere at
both instances needs to be accounted for. Therefore, an identical TEC ramp in both
images, e.g., due to the diurnal variation, will cancel in the interferogram. Examples
of the quantity of delay are given in table 6.2.
Following eq. (6.3.6), for ionospheric effects to have an observable influence in a SAR
interferogram, the dominant wavelength should be less than 100 km, to distinguish
212 Chapter 6: Atmospheric monitoring

them from orbit errors, while the amplitude should be more than 0.36 TECU, to
get a signal of 0.2 phase cycles. Moreover, the anomaly should appear in only one
of the two acquisitions.
The only natural phenomena that could possibly cause such effects are small scale
Traveling Ionospheric Disturbances (TIDs), with a wavelength of tens of km (Spoel-
stra and Yi-Pei, 1995; Spoelstra, 1997). These TIDs are wave effects, mostly prop-
agating from polar regions to regions with lower latitudes.

6.3.3 Spatial scales

The ionosphere can be divided in a number of layers, which have different charac-
teristics. The D-layer (80–100 km) receives only a minor part of the solar radiation,
since most of the energy has been absorbed by higher ionospheric layers, and is
therefore only weakly ionized. After sunset, this layer disappears completely. The
E-layer (100–140 km) has little more ionization, but mostly fades into the F1-layer at
night. The Fl and F2 layers, are located at 140–200, and 200–400 km, respectively
(Afraimovich et al., 1992). TIDs occur mainly in the F-layers.
Medium-scale TIDs have scale lengths of 100–200 km, time scales of 10–20 min,
and result in a 0.5–5% variation in the total electron content. Large-scale TIDs are
relatively uncommon. They have scale lengths of 1000 km, time scales of hours, and
can cause up to 8% variations in the total electron content (Thompson et al., 1986,
p.449). Smith et al. (1950) found that irregularities in the ionization (blobs) occur
at scale sizes of a few kilometers or less. It is not reported, however, how large these
disturbances are as a fraction of the total electron content (Thompson et al., 1986).
Ionospheric gravity waves occur, e.g., when a sudden disturbance in electron density
(e.g., a solar eclipse, waves over a mountain edge, heavy thunderstorms, earthquakes
or rocket launches) triggers an expanding wave front (Calais and Minster, 1995, 1996;
Ho et al., 1998; Calais and Minster, 1998). These waves normally have enormous
wavelengths (hundreds of kilometers), and have gravity as restoring force. Cheng
and Huang (1992) reported waves with a wavelength between 160 and 435 km after
the eruption of Mount Pinatubo in 1991.
 6.3 Ionospheric influence 213

6.3.4 Comparing tropospheric and ionospheric signal

The difference between ionospheric and tropospheric effects on SAR interferometry

is based on two characteristics: the sign of the delay and the shape of the feature.
From eq. (6.3.5), it follows that an increase in the ionospheric electron content results
in a decrease of the observed range, or a phase advance. An area of increased partial
water vapor pressure, see eq. (6.1.7), results in an increase in the observed range,
or a phase delay. However, there are three complications. First, one has to know
during which of the two SAR images the effect occurred: a phase advance in the
first image gives the same result as a phase delay in the second one. Second, one
has to distinguish which area in the interferogram can be regarded as reference area,
where the phase is relatively undisturbed: a localized area with reduced water vapor
density (e.g., a “hole” in a cloud layer) yields a localized relative advance of the
phase. Third, a localized increase in temperature may also result in a relative phase
advance, although this effect usually has a limited magnitude.
The first complication can be resolved by using more than two SAR images. The
different interferometric combinations of images enables the unambiguous identifi-
cation of the image containing anomalies, as long as other error sources such as
temporal decorrelation allow for such an evaluation. As long as single interferomet-
ric pairs are considered, it is not possible to distinguish between phase delay and
advance based on only interferometric data.
The second complication, however, is much harder to circumvent. Just as localized
areas with an increased water vapor content occur frequently, in a relative sense
also areas with a lesser amount of water vapor occur. This can be clearly observed
in, e.g., Meteosat water vapor channel imagery. Therefore, without any additional
information it is not possible to uniquely identify ionospheric effects in interferograms
based on the phase sign of the feature.
The remaining possibility for identification is based on the shape and characteristics
of the feature. The main difficulty in this respect is the limited knowledge about the
spatial characteristics of the ionosphere within ranges of less than 100 km. Although
different devices using dual frequencies have observed the ionosphere, their results
have either been only one-dimensional integrated profiles in time (as is, e.g., possible
with GPS), or large-scale interpolated ionospheric maps that do not have the spatial
resolution needed for InSAR.
Based on these evaluations, it is difficult to uniquely identify ionospheric effects in
SAR interferograms at present, unless strong support from additional sources is avail-
able. It should be noted that further research in this field, using, e.g., dual-frequency
SAR interferometry, could be extremely valuable for its ionospheric imaging capac-
ities.

6.3.5 Reported results

Saito et al. (1998) used GEONET, the permanent GPS array of Japan with a mean
distance between the receivers of 25 km, to map the two-dimensional TEC perturba-
tions. TIDs, with a wavelength of 300–400 km, traveling with a speed of ~150 m/s,
214 Chapter 6: Atmospheric monitoring

were observed. The amount of TEC variation was maximally 0.8 TECU. This
would result in 1.2 cm phase variation in a C-band SAR interferogram. However,
for these wavelengths such variations would result only in a slight curvature of the
interferometric phase.
Gray et al. (2000) observed ionospheric wave effects by examining the azimuth offset
vectors during the coregistration of two SAR images over the Antarctic (81.3°S). This
resulted in a wave-pattern with wavelengths of ~5 km which could be explained by
a variation of 0.12 TECU (less than 2 mm zenith delay difference) during one of
the SAR acquisitions. The variable electron content caused an erroneous Doppler
frequency, resulting in an offset of the coverage of resolution elements in azimuth.
In the interferometric phase these effects were not visible, likely due to their limited
magnitude and the phase contribution of topography, deformation, and tropospheric
signal.
Massonnet and Feigl (1995a) reported ionospheric effects in interferograms over
Landers, California (34.5°N). They identified a kidney-shaped anomaly by pair-
wise logic, to occur in a single SAR acquisition. This anomaly covered an area of
10 × 30 km, oriented approximately north-south, with a maximum amplitude of
~3 cm (approximately one fringe). The main reason for assuming an ionospheric
driving mechanism is that the anomaly shows a phase advance, instead of a phase
delay. Massonnet et al. (1995) show another example of presumed ionospheric ori-
gin over Etna volcano, Italy (37.7°N). Here a triangle shaped anomaly with sides of
~4 km and a phase advance of approximately one fringe is observed.
A number of arguments can be listed that oppose this argumentation. First, the
magnitude of the effects is probably too large. From eq. (6.3.6) we find that a
rather large local increase in electron content of 2 TECU is needed to explain 3 cm
magnitudes. For example, TIDs observed by Saito et al. (1998) with a wavelength
of 300–400 km show a maximum effect of 0.8 TECU. It seems reasonable to assume
that amplitudes decrease for decreasing wavelength, following some sort of power-
law behavior. This argument is supported by the observations by Gray et al. (2000),
which show that wave-like features with a wavelength of ~5 km in auroral zones
have a magnitude of 0.12 TECU. Second, the size, shape, and isolated nature of the
two observed anomalies is peculiar. Known ionospheric anomalies behave like wave
effects. Even ionospheric perturbations observed after a Space Shuttle ascent—a
very discrete phenomenon—show wave effects, with a magnitude of 0.04–0.2 TECU
(Calais and Minster, 1996). The wave effects observed by Gray et al. (2000) and
Joughin et al. (1996) had wavelengths of several km’s, but extending over hundreds of
km’s. The isolated nature of the anomalies observed by Massonnet and Feigl (1995a)
and Massonnet et al. (1995) does not correspond with any form of anomaly found
in literature. Finally, although this does not necessarily prove that such anomalies
cannot exist, the lack of a possible driving mechanism for the presumed ionospheric
perturbations makes this hypothesis less likely.
An alternative hypothesis to explain the two examples can be easily found by con-
sidering a tropospheric explanation. For example, a relative phase advance can
result from a cloud, or (better) water vapor, cover with “holes” in it. A layer of
homogeneous stratus or stratocumulus cloud could easily result in a fringe of phase
 6.4 Water vapor mapping 215

delay. In fact, if such a layer would cover the entire interferogram, it would not even
be observed over a flat area. A localized area with less water vapor and/or cloud
droplets would result in a relative phase advance as well. Finally it needs to be
noted that a localized increase in temperature will also result in a phase advance,
see section. 6.2.1, although the magnitude of these effects is usually limited to a
couple of millimeters.

6.3.6 Discussion

Based on (i) the theoretical considerations discussed in this section and (ii) the
experimental results based on GPS and InSAR observations, we conclude that until
today the influence of ionospheric disturbances on repeat-pass spaceborne C-band
radar interferograms is at the least controversial, especially for mid-latitudes. The
results of Gray et al. (2000) give a first proof of ionospheric effects in auroral zones,
even though the derived magnitudes were not clearly identified and validated by
analyzing the interferometric phase. Therefore, we currently adapt the hypothesis
that ionospheric effects may result in long wavelength gradients/curves over a single
SAR image, but will not noticeably affect phase variations at scales less than ~50 km.

6.4 Water vapor mapping

Three SAR interferograms, obtained over the Netherlands (fig. 6.5), show several
prominent and representative features. The SAR data were acquired by the ERS-
1 and ERS-2 satellites, operating in a one-day interval mode, ensuring sufficiently
coherent images over land areas. The principle ground resolution of 4 x 20 m is
spatially averaged to 160 × 160 m, reducing the delay standard error from a couple
of millimeters to below 1 mm. Maps of the water vapor distribution associated with
a precipitating cloud, a partly precipitating cold front, and horizontal convective
rolls reveal quantitative measures that are not observed with conventional methods,
and suggest that such radar observations can be used for forecasting and for studying
atmospheric dynamics. These result have been reported by Hanssen et al. (1999).

6.4.1 Precipitating cumulonimbus

The signature of a precipitating cumulonimbus cloud is shown as a localized delay

difference in the SAR line of sight, caused by the cloud and the saturated sub-cloud
layer see fig. 6.6A. Using the method described in Davis et al. (1985); Bevis et al.
(1992) and a simple cosine mapping function, see eq. (6.1.4), the delay differences are
mapped to differences in zenith integrated precipitable water the vertically
integrated water vapor liquid equivalent. The positive sign of the delay indicates that
the feature appeared in the first of the two combined SAR images. This inference is
verified by the weather radar reflectivity signal, expressed as rain rate, on the first
day (fig. 6.6B). During the second SAR acquisition, no precipitation was observed
in the weather radar data. Surface observations and radiosonde data estimated that
the cloud base height was 800 m; weather radar data indicated that cloud tops were
as high as 4.5 km.
216 Chapter 6: Atmospheric monitoring

The shape and magnitude of the anomaly in the interferogram correspond closely to
the weather radar echo. Phase variations over areas with no precipitation suggest
that water vapor gradients are seen in the interferogram that are not observable by
the weather radar. The signal over the cloud has a mean value of 1.3 mm,
which is reasonably close to the 0.9 mm value inferred from a radiosonde profile
considering it was launched about 90 km away and 4 hours prior to the SAR overpass.
The interferometric technique can measure only relative changes in water vapor,
nonetheless the data reveal that this storm system has approximately liters
more water compared to the surrounding air. For this system to reach equilibrium,
this excess water would have to be redistributed over a larger area or precipitate out.
Such numerical constraints form important information for forecasting purposes.

6.4.2 Partly precipitating cold front

A cold front, propagating from the northwest, is shown in fig. 6.7A. The interfero-
gram reveals a narrow diagonal band of enhanced water vapor, suggesting the pres-
ence of a narrow cold-frontal rain band (Parsons, 1992). The weather radar data
(fig. 6.7B), however, show that only parts of the band are precipitating. Surface
winds (fig. 6.7B) indicate some convergence at the cold front. The frontal circula-
tion may have produced a localized maximum in magnitude or depth of water vapor
at the convergence zone, visible as the yellow band in the interferogram. Studies
using in situ aircraft and radiosonde observations suggest that regions of enhanced
moisture may occur at low-level convergence zones (Simpson, 1994). The sickle-
shaped area visible at the southern end of IJssel Lake likely reflects an evaporatively
 6.4 Water vapor mapping 217

cooled surface outflow (gust front) generated by the rain showers.

The post-frontal relatively cold and humid air produces a delay of about 28 mm
compared to the pre-frontal, relatively warm air. This increase is partly caused
by an overall temperature difference of approximately 3°C associated with the two
air masses. The frontal band is 3–5 km wide and has an average increase in delay
between 11 and 64 mm. Using a cold front depth of 3 km, this delay could either be
produced by a 3 hPa increase in water vapor or by an unreasonably high temperature
decrease (10°C) at the frontal boundary. Additionally, the contribution of liquid
water in a cumulus cloud line would produce a maximum delay difference of less
than ~2 mm (Kursinski et al., 1997). Thus water vapor, rather than temperature
or liquid water, is likely the primary mechanism causing the increased delay at the
frontal boundary.
The areas with delay differences of more than 30 mm correspond to precipitation
regions. The sudden increase of water vapor content in the sub-cloud layer, caused by
partial evaporation of the precipitation, accounts for this increase. The liquid cloud
and rain particles cannot produce the enhanced delay, as the precipitating clouds
would have to be over 21 km deep to account for the maximum observed delay of
64 mm. The observed features in the interferogram illustrate how localized areas
with an increased water vapor content may give an indication for the occurrence and
development of precipitation.
218 Chapter 6: Atmospheric monitoring

6.4.3 Boundary layer rolls

The interferogram in fig. 6.8 illustrates the frequently-occurring phenomenon of hor-

izontal convective rolls, a form of boundary-layer convection in which there are
counter-rotating horizontal vortices with axes aligned along the mean boundary-
layer wind direction (LeMone, 1973; Brown, 1980). Some evidence that the linear
features of the interferogram depict rolls rather than atmospheric waves is that the
bands are oriented along the wind direction, as shown by the surface station data
overlaid on the interferogram (fig. 6.7A). Moreover, radiometer satellite imagery
 6.4 Water vapor mapping 219

(fig. 6.7B) to the south of the interferogram location shows some evidence of cloud
streets oriented with the interferogram bands. Sufficient moisture in the presence of
rolls often produces cloud streets atop roll updraft branches (LeMone and Pennell,
1976). Furthermore, a local radiosonde on 4 April 1996 at 12:00 UTC indicated a
boundary layer depth of 800 m. This result, combined with the band spacing of 2–
3 km, gives an aspect ratio of 2.5–2.9, consistent with the presence of rolls (Kuettner,
1971).
Roll updrafts are warmer and moister than roll downdraft branches (Weckwerth
et al., 1997). Moisture variability is likely dominant in the observed 6–10 mm delay
differences, because vapor pressure variations associated with rolls (3 hPa) would
produce a delay of 13 mm, while temperature variations of 0.5°C would produce
less than 1 mm delay. The observations show how streets of water vapor can be
identified even when cloud streets are not visible. In fact, no clouds were observed
from the weather stations in the interferogram area. Knowledge of these systematic
patterns of water vapor can aid in forecasting weather behavior.

6.4.4 Discussion

Radar interferometry can provide worldwide coverage of mesoscale atmospheric phe-

nomena over land and ice, and easily monitor more exotic features such as grav-
220 Chapter 6: Atmospheric monitoring

ity waves accompanying fronts. Readily attainable integrated precipitable water

amounts over large areas may make pinpoint weather forecasting a possibility. The
delay maps could eventually be used by bench forecasters or serve as an additional
constraint in variational data assimilation models. Contemporary satellites, how-
ever, have an orbital repeat period that is far from ideal for operational applications,
resulting in loss of coherent phase information over many areas. Ideally, SAR data
should be acquired with multiple short temporal spacings in order to maintain high
coherence and permit a large number of interferogram pairs to be generated. Multi-
ple observations give not only better precision in time, but keep coherence and hence
data quality high. Proper use of airborne and spaceborne SAR systems with short
repeat periods could lead to much greater accuracy in meteorological understanding
and forecasting.

6.5 Combining amplitude and phase information

In the previous section it was shown how observations from radar interferometry can
be used to extract information on the moisture distribution in the boundary layer.
Here we demonstrate how the same radar observations can also be used for wind
field measurements, which enables an integrated analysis of wind and moisture pa-
rameters from a single sensor. Since mesoscale surface winds modulate fluxes of mo-
mentum, heat, and moisture—the driving forces of atmospheric circulation—there
is close correspondence between the wind field and the transport and distribution
of moisture. Examples include, e.g., cumulus convection and atmospheric dynam-
ics near fronts. Improved understanding of both spatial and temporal behavior of
air masses is important for parameterizations in weather forecasting and climate
models.

Generally, basic meteorological parameters such as wind velocity and water vapor
distribution are not measured on km-scales (see, e.g., Stoffelen (1998) for wind mea-
surements and Susskind et al. (1984) for temperature and humidity information).
A related problem is that wind and moisture observations are usually acquired by
different sensors, that don’t necessarily coincide in time. For the analysis of causal
connections between both types of parameters on scales between 100 m and 10 km,
temporal coincidence is imperative due to their spatial and temporal variability.

Here we elaborate on two case studies exploiting both the radar backscatter in-
tensity and phase information. This approach provides a more thorough examina-
tion of previous findings, which were based on phase information only. Using two
pairs of interferometric SAR observations the moisture distribution over land areas
is retrieved from the interferometric phase, whereas wind information is retrieved
from the backscattered radar energy over water areas. The case studies constitute
a comprehensive quantitative examination of vortices associated with a system of
boundary layer rolls and with a rain band in relation to a cold front. This study has
been reported by Hanssen et al. (2000b).
 6.5 Combining amplitude and phase information 221

6.5.1 Methodology and data analysis

Spaceborne SAR provides fine-resolution radar images over broad areas. The reso-
lution of the ERS images analyzed here is ~20 m over a swath of 100 km in width.
The radar data comprise a grid of complex vectors, one for every pixel. The length
of the vector is a measure for the backscatter intensity and yields information on
the roughness of the imaged surface. Over water surfaces, the backscatter intensity
is dominated by Bragg scattering (Curlander and McDonough, 1991). Contrasts
in surface roughness, quantified in the normalized radar cross section can be
interpreted as differences in mainly wind speed and direction (Stoffelen, 1998). Sea
state or heavy rain are further atmospheric phenomena that may influence the in-
tensity. SAR images can detail small-scale structures over seas such as boundary
layer rolls (Alpers and Brümmer, 1994). Here we used a C-band model, CMOD4, to
compute the wind speed from the normalized radar cross section and the incidence
angle of the radar waves (Stoffelen, 1998). This model correlates backscatter inten-
sity with the wind at 10 m height. Note that other models correlate to a so-called
neutral equivalent wind, but in that case accurate ancillary information is needed
on atmospheric stratification (Stoffelen, 1998).
Here we eliminate phase signal due to geometric path length differences using an
elevation model (Massonnet et al., 1993). As the interferometric pairs are acquired
with a time interval of only 24 hours, coherent phase information is obtained in which
no surface deformation occurred. This methodology ensures that the interferometric
phase is only due to atmospheric propagation delay.
The delay of the radar signal is caused by an integration over the refractive index
of the propagation medium, along the line of sight. Horizontal and vertical hetero-
geneities in refractive index are influenced by the spatial distribution of water vapor,
pressure, temperature, liquid water, and electron content. Yet, over distances less
than ~50 km, the main signal in the interferogram is due to water vapor, albeit
temperature and liquid water can add some additional millimeters of delay. Using
surface temperature observations, the integrated water vapor signal can be con-
verted to precipitable water—its liquid equivalent, when assuming a fixed vertical
temperature profile (Bevis et al., 1992).

6.5.2 Results

Case study 1: Precipitation and convection

An ERS-2 SAR image (orbit 2388, frame 1053) was acquired during a cold front
passage over the Netherlands, on 4 October 1995. The interferometric combination
with an ERS-1 SAR image (orbit 22061), acquired 24 hours earlier, in the absence
of significant weather in our target area, enabled the construction of a radar inter-
ferogram, see fig. 6.9. The figure shows coherent phase information over all land
areas. Over the water area the intensity image of the SAR acquisition of 4 October
is shown. The diagonal line in the lower-right corner of the image is associated with
the location of the cold front, and the curved anomaly in the center of the image,
south of the lake, can be associated with a gust front. Weather radar observations
222 Chapter 6: Atmospheric monitoring
 6.5 Combining amplitude and phase information 223

show strong precipitation in this region, see fig. 6.7B.

Figure 6.10 shows a small section of this interferogram, centered over Veluwe Lake.
The footprint of the storm is moving over the area from left to right. Over the
land area, the increased humidity associated with the convective clouds and rain is
measured as an increase in signal delay of approximately 30 mm. At this position,
P4, weather radar observations indicated rain rates up to 10 mm/hr. In front of
the storm (P5) the interferogram shows approximately uniform color, indicating not
much variation in moisture. Behind the storm front (P3), variations in moisture are
in the order of 8–12 mm, which can be associated with moisture variations due to,
e.g., convective cells. The small circular anomalies near P5 are related to processing
errors over single pixels.
The roughness of the water surface is expressed by the normalized radar cross sec-
tion, and is well correlated with the near-surface wind conditions (Stoffelen,
1998). It is likely that the signatures on the water surface are mostly due to wind
effects, as the signature of the rain would cause fine-scale structures. Moreover, the
rain rate is probably too low and the wind speed likely too high for there to be a
rain effect by impact on the water surface. A cross-section of fig. 6.10A is shown
in fig. 6.10B. Collocated with the rain band, an increase in of approximately
18 dB can be observed, at position P4. To the east of the storm front (P5) the
water surface is relatively smooth. Moving towards the west, from P4 to P2, the
roughness first decreases approximately 8 dB (P3), to increase approximately 4 dB
at P2. This nearly symmetrical wind shear appears to be closely associated with the
storm front. Assuming a wind direction somewhere between upwind and crosswind
(the horizontal wind components in the radar look direction and perpendicular to it,
respectively, see fig. 6.10A) the wind velocities following from CMOD4 are
in front of the storm (P5), increasing up to in the storm (P4), within a
spatial distance of 300 m.
At P4, there is a surplus of water vapor with respect to the surroundings, resulting in
an excess delay of 3.6 cm, which corresponds to 5.3 ± 0.5 mm integrated precipitable
water. The width of the main rain band is 2 km.
From wind observations and common circulation patterns in high-latitude convective
systems, one can infer that the radar look direction is crosswind to the flow at
location P5 in fig. 6.10A, and parallel to the rain band. Behind the rain band, P1-
PS, the wind is more along the radar look direction and probably veering. More to
the northeast of the rain band surface measurements indicate that the wind has also
a component along the radar look direction. This means that besides wind speed
changes, there are also wind direction changes that may be relevant for the signature
in the radar image (P1–P5). The radar is most sensitive to wind direction changes
for wind directions at 45, 135, 225, and 315 with respect to the radar look direction,
whereas there is no sensitivity at 0, 90, 180. and 270 degrees.
From these observations, we find a circulation pattern connected to this system
termed “optimal shear” by Parsons (1992), depicted in fig. 6.10C. In our case the
complete system moves to the east, such that the vertical shear in front of the storm
consists of an area with no wind at the surface and probably SW-winds aloft.
224 Chapter 6: Atmospheric monitoring
 6.5 Combining amplitude and phase information 225

Case study 2: Boundary layer rolls

In section 6.4.3 it was discussed how during an ERS-1 acquisition on 4 April 1996,
10:34 UTC (orbit 24688, frame 2547), convective rolls occurred in the boundary
layer. In the interferogram with an ERS-2 acquisition on 5 April 1996 (orbit 5015),
see fig. 6.11, these rolls are clearly visible over the land areas. Since no clouds were
observed at the surface stations in the area, the delay variation is a clear air effect.
Several studies have indicated that the ascending parts of the rolls transport warm,
moist air at the surface upward, while the descending parts transport dry air orig-
inating within the inversion downward, see, e.g., LeMone (1973). Weckwerth et al.
(1997) observed variations in the water vapor mixing ratio of 1.5–2.5 g/kg associated
with rolls, whereas temperature variations were only 0.5 K. Mapped to water vapor
pressure, these moisture variations are 2.4–4 hPa. The sensitivity of the delay to a
1 hPa change in water vapor pressure is in this situation at least 3.5 times as large as
to a 1 K change in temperature. The delay variations of up to 10 mm are therefore
primarily caused by water vapor differences in the updraft and downdraft branches
of the rolls. Not more than 5% of the observed delay can be explained by the tem-
perature variations in the rolls. The rolls are approximately parallel to the wind
direction, and the wind speed observed at several surface stations is
Over the water areas in IJssel lake, wind streaks are clearly visible, oriented in
the same direction as the rolls over land areas. Although streaks caused by rolls
are frequently observed in SAR images, see Alpers and Brümmer (1994), Mourad
and Walter (1996), the location of the streaks with respect to the downdraft and
updraft branches has been uncertain. Spectral analysis reveals a streak spacing of
approximately 2 km. Over land, the spacing between the bands is also approximately
2–3 km. A radiosonde in the area, launched 90 minutes after the SAR acquisition,
indicated a boundary layer depth of 800 m, which gives an aspect ratio of 2.5–2.9,
characteristic for rolls (Atkinson and Zhang, 1996).
For wind speed estimation with a accuracy, an averaging area of
is needed (Lehner et al., 1998). Using square areas of 10 x 10 km, average wind
speeds of were found. For the analysis of the wind speed differences
between the streaks, square areas of this size cannot be used. Instead, we applied
a normalized Radon transform to two areas of 11 x 11 km and used a profile per-
pendicular to the streaks to estimate the average differences in (Hanssen, 1994).
This method reduces speckle effects significantly and enables direction dependent
amplitude estimation. Variations over a range of 0.4–0.5 dB are observed, corre-
sponding with wind speed variations of or wind direction variations of up to
45 degrees (i.e., the sensitivity to lateral wind velocity changes is about four times
the sensitivity to transverse velocity variations).
We note the change in characteristics of the boundary layer rolls as they are ad-
vected over the cold water surface, where the fluxes of momentum and heat are
expectedly quite different from those over land areas that are heated by the sun.
The combination of measurements shown in this paper is ideally suited to show the
transformation of air flow after advection over a land-water boundary.
The location of the streaks aligns approximately with the part of the rolls that have
226 Chapter 6: Atmospheric monitoring

less delay, which correspond to the relatively dry and descending air. This is in
correspondence with classical theory, which expects the maximum wind speed in a
roll in the downdraft area (Alpers and Brümmer, 1994; Atkinson and Zhang, 1996).
Mid-latitude cases, without cloud and with important dynamical and thermal insta-
bility, such as presented here, are uniquely documented by the SAR methodology
presented here.

6.5.3 Discussion

The possibility to simultaneously combine both radar intensity observations over

water and phase observations over land surfaces enables specific studies on the in-
terrelation between moisture distribution and wind fields. The fine spatial resolution
is particularly suitable for the study of mesoscale shallow convection in the plane-
tary boundary layer in coastal areas or areas with large lakes to describe for example
the transformation of air across the land-water interface. Although the phase delay
measurements are always a summation of the delay during two acquisitions, specific
patterns can often be recognized and analyzed.

6.6 Validation radar interferometry using Meteosat radiometry

Validation of the radar interferometric results can be difficult since conventional

imaging radiometers do not provide quantitative measures for water vapor content
over the entire tropospheric column and have poor spatial resolution (Weldon and
Holmes, 1991). Moreover, comparable quantitative data such as signal delay ob-
served by Global Positioning System (GPS) receivers are only available as time
series at a fixed location. In this section, the technique of InSAR integrated re-
fractivity mapping is validated for a specific atmospheric situation with relatively
transparent air. It is investigated whether it is possible, for this specific situation,
to obtain a parameterization of radiometer brightness temperatures as a function of
precipitable water vapor, using Meteosat [water vapor (WV) channel] ob-
servations and GPS time series. After establishing such a relationship for a single
point, it is extended spatially to derive precipitable water vapor that can be used
for validating the SAR interferogram.
In section 6.6.1 the basic characteristics of the Meteosat observations are described.
Section 6.6.2 describes the methodology to derive precipitable water vapor estimates,
and presents the main results of this study. Discussion and conclusions make up
section 6.6.3. These results are reported by Hanssen et al. (2001).

6.6.1 Meteosat water vapor channel

The Meteosat WV channel, centered around is used in operational meteo-

rology to observe the development of structures of upper-tropospheric water vapor,
which carry the signatures of atmospheric conditions. Especially subsidence inver-
sions, that is, downward vertical transport of dry air behind a frontal zone, are
clearly visible in the WV images. Due to the strong absorption by water vapor at
this wavelength, the observed brightness temperatures usually originate from tro-
 6.6 Validation radar interferometry using Meteosat radiometry 227

pospheric layers above 3 km (Weldon and Holmes, 1991). Therefore, quantitative

analysis is restricted to upper-tropospheric water vapor as described by Schmetz
et al. (1995). Unfortunately, however, the concentration of water vapor is highest
near the earth’s surface, where relatively high pressure and temperature allow the
air to contain more water vapor. Therefore, in general it is not possible to make one
228 Chapter 6: Atmospheric monitoring

unique parameterization of WV channel brightness temperatures into precipitable

water vapor. In section 6.6.2 it is discussed whether GPS tropospheric delay time
series are suitable for establishing a tailor-made parameterization to obtain Meteosat
spatial precipitable water vapor values for the special case of a subsidence inversion,
at a local resolution of 5x9 km.

6.6.2 Methodology and results

SAR interferogram analysis

An interferogram. covering a strip of 100 x 200 km, has been formed using ERS
SAR acquisitions on 26 and 27 March 1996, 21:41:05 UTC (22:41:05 local time).
The orbital separation parallel to the look direction, see fig. 6.1, is approximately
32 m, which makes the configuration moderately sensitive to topographic height
differences. An a priori reference elevation model is used to correct for the topo-
graphic phase in the interferogram (TDN/MD, 1997). Some additional corrections
for tilts in the length and width direction are applied, and water surfaces are masked.
The resulting differential interferogram is shown in fig. 6.12. The phase values are
converted to zenith wet delays, using eqs. (6.1.3) and (6.1.4), and consecutively to
differential precipitable water vapor using eq. (6.6.3), where we use the term “differ-
ential” to indicate the difference between the two states of the atmosphere during
the SAR acquisitions.
A strong, large-scale, gradient is clearly visible in the south of the interferogram,
aligned approximately perpendicular to the radar flight direction. The interpretation
of this phenomenon is a key topic of this study. Previous studies have shown that
phase variation perpendicular to the flight direction might be caused by oscillator
drift errors in the onboard reference clock (Massonnet and Vadon, 1995). A possible
way to examine this possibility would be the calculation of a long swath of connected
SAR images, but this is outside the scope of this study. Here it is assumed that the
phase variation is caused by atmospheric water vapor only. The likelihood of this
assumption can be tested using the combined analysis with Meteosat and GPS.
Note that, apart from the gradient in the south of the interferogram, wave phenom-
ena also can be observed in the interferogram, see the lower left-hand part of fig. 6.12.
These are possibly due to low-level moisture variations of which the structure seems
to indicate the presence of boundary layer rolls.
Wet delay differences can be related to integrated precipitable water (I) values,
the liquid equivalent of the integrated water vapor:

Using the density of liquid water, we find that 1 mm integrated precipitable water is
equal to integrated water vapor. The conversion factor is approximated
using surface temperatures for the image acquired at defined by Askne and
Nordius (1987). After first models for were derived by Davis et al. (1985) and
Bevis et al. (1994), a polynomial model was developed to approximate for De
Bilt, the reference site in the Netherlands, using 1461 radiosonde profiles (Emardson
6.6 Validation radar interferometry using Meteosat radiometry 229
230 Chapter 6: Atmospheric monitoring

and Derks, 2000):

where is the mean annual surface temperature for the location, and is the
day number of the year. The used coefficients are
and (Emardson
and Derks, 2000). Using surface temperatures provided by 18 stations in the area,
we find that and for 26 and 27
March respectively. Applying one factor for both days, is
justified, since the fractional error in II is one order of magnitude smaller than the
fractional error in the delay measurement. We can rewrite eq. (6.6.1) as

where we define as the double difference of integrated precipitable water

in space and time. Note that a single, isolated anomaly will have a different sign
in the interferogram, depending on whether it appeared in the first or the second
acquisition. In that case, one of the single spatial differences can be regarded
as nearly zero, in which case we can interpret the double difference as a single spatial
difference during the other acquisition.
Meteosat WV channel analysis
On 27 March 1996 a strong subsidence inversion passed the Netherlands from north
to south, which was clearly visible in the WV channel images (fig. 6.13). The sub-
sidence inversion can be identified by the dark band in the image, corresponding
to relatively high brightness temperatures. Because the descending air in the sub-
sidence inversion is rather dry, the absorption (and emission) of radiation is low
and, therefore, the air is relatively transparent. This enables radiation from lower
(warmer) layers to contribute to the signal, which results in high apparent brightness
temperatures. The center of the subsidence inversion moves in one day from about
53°N to 48°N, as apparent from fig. 6.13.
Assuming that the brightness temperature variations in the Meteosat WV images re-
flect the variations of the total water vapor column, we use GPS-derived precipitable
water vapor observations to parameterize the brightness temperature. The reference
to the WV channel are collocated measurements from GPS ground station Kootwijk,
at 52.17°N, 5.80°E, see fig. 6.12. From the GPS measurements reliable absolute
values for precipitable water vapor are derived, using the standard methodology de-
scribed in Bevis et al. (1992, 1994). The accuracy of the GPS-derived precipitable
water vapor is approximately 2 mm. For 27 hourly observations of GPS precipitable
water vapor and Meteosat signal intensity, shown as scatterplot in fig. 6.14, the
correlation can be parameterized as
6.6 Validation radar interferometry using Meteosat radiometry 231
232 Chapter 6: Atmospheric monitoring

with I in and the signal intensity of the WV channel in number of counts.

The rms value of the difference between model and observations is
In fig. 6.15 the change in I over Kootwijk is shown derived from both GPS and
Meteosat. We find that the rms of is accurate enough for our analysis.
The passage of the subsidence inversion from north to south is clearly visible. Values
range from 1 to
Using the parameterization in eq. (6.6.4), precipitable water vapor can be derived
from the Meteosat WV channel image. The parameterization yields accurate val-
ues for the area near Kootwijk (location indicated in fig.6.12), but with increasing
distance the accuracy is likely to decrease.
For the comparison with the SAR interferogram an elongated area is selected (in-
dicated by the arrow in fig. 6.13), which ranges from 51.38°N, 5.93°E to 52.72°N,
5.43°E. The area includes the GPS station Kootwijk Observatory for Satellite Geodesy
(KOSG), to optimize the validity of our parameterization, and is parallel to the
 6.6 Validation radar interferometry using Meteosat radiometry 233

ground track of ERS. The subsidence inversion band is nearly perpendicular to the
profile and moves approximately parallel to it. As a result the analyzed precipitable
water vapor signal is dominated by only one atmospheric process—the subsidence
inversion. Therefore, the parameterization is assumed to be sufficiently accurate for
the selected profile.
All pixels that contain high-altitude clouds (temperature below – 20° C) are excluded
from the analysis, to increase the reliability of the parameterization. In a cloud,
the fraction of total water that is clustered in particles is generally small, but the
absorption of radiation by these particles is much larger than the absorption by
the water vapor in the cloud. Therefore, the correlation between the Meteosat
WV channel observations and the correct amount of precipitable water vapor will
deteriorate if clouds are present.
To obtain similar quantities as in the interferogram the extracted profiles corre-
sponding with 22:00 UTC 26 March and 22:00 UTC 27 March are differenced. The
resulting values are indicated by the triangles in fig. 6.17, where the position of
the pixels along the profile is expressed by their latitude. From the Meteosat WV
images in fig. 6.13 it appears that at 26 March, 22:00 UTC, the selected area was
in the front part of the subsidence inversion, which results in a decreasing amount
of precipitable water vapor with latitude. At 27 March, 22:00 UTC, the selected
area was in the back part, resulting in an increasing amount of precipitable water
vapor with latitude. As a consequence of these opposite trends, the quantities in
fig. 6.17—which are formed by subtracting the values at 27 March from the values
at 26 March—have an amplified north-south gradient. In the following these results
will be compared with the results from SAR interferometry.
Intercomparison
The evaluation of the water vapor observations from the Meteosat WV channel and
the radar interferogram is subject to 5 degrees of freedom.

1. The interferogram shows relative delay differences. Therefore the analyzed

profile has an arbitrary bias when compared with the absolute values of the
Meteosat profile.
2. Due to the oblique viewing geometry of the radar (23° from zenith) the profile
will be shifted some kilometers parallel to the west, see fig. 6.16. Since the
atmospheric situation during the two acquisitions was nearly symmetric (i.e.,
perpendicular to the profile), the effect of this lateral shift is negligible.
3. The inaccuracy in the satellite orbits might lead to a small tilt in the profile.
Here we assume that this tilt is sufficiently eliminated by using a number of
reference points during preprocessing.
4. The Meteosat positioning accuracy is approximately 0.5 pixels, corresponding
to a 4.5-km uncertainty in north-south direction.
5. Due to the geostationary position over the equator, there is an additional
shift D in north-south direction when the majority of the water vapor is at
234 Chapter 6: Atmospheric monitoring

a height see fig. 6.16. In that case, the information will appear to be
shifted northward in the image. For the location analyzed here, this shift is
approximately Note that the time difference between the SAR
acquisition and the Meteosat scan of the latitudes of the Netherlands (13 min)
might also account for a slight additional shift.

Estimates of the a priori standard deviation of both the Meteosat and InSAR pre-
cipitable water vapor estimates are needed to evaluate the correlation between the
two profiles. For the Meteosat observations, is assumed to be uncorrelated
and equal for every observation. In section 6.6.2, an rms value of was
derived from a comparison with GPS observations. The analyzed profile values are
differences between corresponding observations for two consecutive days. Assuming
zero covariance between the two days, standard error propagation yields the a priori
value for
The expected variance for the InSAR observations is assumed to be uncorrelated
between adjacent values and equal for every observation. In section 6.1.2, the delay
standard deviation was derived. Using eq. (6.6.3), simple
error propagation yields an a priori value for the integrated water vapor observations
of
A goodness-of-fit parameter between the InSAR and the Meteosat observations can
be optimized by adjusting the 5 degrees of freedom mentioned above, or by adjusting
the a priori variances of both observations. Here, a combination of both approaches
is suggested. Parameters 2–4 are assumed to be sufficiently well approximated. A
northward shift of the Meteosat observations (parameter 5) of 5 km is used as a
first-order approximation, derived from an average height of the dominant water
vapor signal in the WV channel. The additional Meteosat profile shift, the bias of
the InSAR profile (parameter 1), and the scaling factor for the variances are now
estimated by deriving the minimal reduced chi-square value, of the two profiles
 6.6 Validation radar interferometry using Meteosat radiometry 235

as a function of the bias of the InSAR profile (Bevington and Robinson, 1992):

The number of measurements is denoted by are the differences between the

Meteosat and the InSAR profile values, and are the a priori standard deviations
of the differences, that is, Here we find that
for all Note that the contribution of is one order of
magnitude larger than
The values are expressed as a function of the InSAR profile bias and the horizontal
shift of the Meteosat profile. The search window of the bias has been confined
between –3 and whereas the horizontal shift is confined between –1.1
and 1.1 km. The minimum value, is found for a bias of and a
horizontal shift of –1.1 km. Figure 6.17 shows the original position of the InSAR
profile, indicated by the thin line, and the new position, indicated by the bold line.
The original position of the Meteosat profile values is indicated by the triangles,
while the new positions are indicated by the squares. A 95% confidence interval for
the difference of the two data sources is indicated by the inner error bars. Note that
the error bars are drawn at the Meteosat positions, although they include a small
component of the InSAR profile as well.
From this evaluation it is clear that the adjustment of the free parameters is not
sufficient to reach a satisfying comparison. The estimated variance can be ap-
proximated by (Bevington and Robinson, 1992)

if all variances are equal. If both profiles describe the same physical process, the
estimated variance should agree well with the a priori variance and the value of the
reduced chi-squared should be approximately unity. In order to reach this situation,
the a priori standard deviations of the difference need to be scaled by a factor
This implies that the a priori standard deviation of the difference was
too optimistic. An a posteriori standard deviation of 1.23 mm is found, indicated in
fig. 6.17 by the outer error bars.

6.6.3 Discussion

For a specific atmospheric situation in March 1996. precipitable water vapor obtained
from SAR interferometry is validated by GPS time delay analysis combined with
Meteosat WV channel observations. The interferometric phase observations
are converted to relative signal delay observations and consecutively processed to
precipitable water vapor. For a point location, Meteosat brightness temperature time
series are converted to precipitable water vapor using a parameterization obtained
from GPS wet signal delay observations. Applying this parameterization spatially
for a profile of brightness temperatures in two WV channel images, acquired at
nearly the same time as the two SAR images, enables a direct comparison between
the two sources.
236 Chapter 6: Atmospheric monitoring

The results show that the phase gradient observed in the SAR interferogram is
fully accounted for by a subsidence inversion that moved over the interferogram
area during the two SAR acquisitions. The subsidence inversion resulted in tem-
poral precipitable water vapor variations over a range of approximately
as observed by GPS and Meteosat. Accounting for the relative character of the
InSAR observations, and the positioning uncertainty of the Meteosat images, both
datasets describe the same phenomenon in a minimal reduced chi-squared sense with
standard deviation. As the range of the signal encompasses approxi-
mately signal to noise ratio values are sufficient to identify the same signal
in both data sources. This supports the statement that radar interferometry can be
used to derive the spatial variations in precipitable water vapor.
Although the parameterization used in eq. (6.6.4) appears to be sufficient to explain
the observations, it needs to be stressed that the use of this method is feasible only for
a limited area around the GPS receiver. Moreover, in situations with severe cloud
cover, the conversion from brightness temperatures to precipitable water vapor is
not likely to succeed, due to insufficient penetration caused by absorption. Future
applications of the technique need to be focused on consecutive acquisitions of SAR
 6.7 Spatial moisture distribution during CLARA 96 237

images to obtain “cascade” series of interferograms. This approach can result in

resolving the ambiguity between the atmospheric states during the acquisitions.

6.7 Spatial moisture distribution during CLARA 96

In this study, published by Hanssen and Weckwerth (1999), a specific interferogram

is analyzed, acquired at 23 and 24 April 1996 over the southwestern part of the
Netherlands, see fig. 6.18. The two SAR acquisitions coincide with the “Clouds
and Radiation” (CLARA) experiment, which is aimed at a better understanding
of the interaction between clouds and radiation by improving the measurement of
cloud properties (van Lammeren et al., 1997; Feijt, 2000). Since the development
of clouds is highly influenced by the water vapor distribution, the SAR data can
be used to estimate the precipitable water content of clouds relative to the ambient
environment. The high resolution of the data reveals regions of enhanced moisture
leading to cloud formation and rain.
Similar space-geodetic techniques, such as GPS, are influenced in nearly the same
way as SAR. Therefore, these data are used to describe the tropospheric situation
during the experiment and identify specific features in the interferogram, in com-
bination with an extensive array of supporting instrumentation. A statistical eval-
uation of the dataset is presented using two-dimensional structure functions, and
conclusions on the potential and limits of radar interferometry as a meteorological
technique are discussed.

6.7.1 Methodology

At 23 April 1996, 10:38:07 UTC (12:38:07 local time), the ERS-1 SAR acquired
frame 2565 during orbit 24960, covering an area of 100 x 100 km. Exactly 24 hours
later, the twin instrument onboard ERS-2, orbit 5287, acquired the same part of
the earth’s surface, from a nearly identical position as ERS-1 one day before. Both
satellite positions differ only 38 m in the oblique look direction of the radar, and
78 m perpendicular to the look direction.
Using a reference elevation model for the test area a differential interferogram is
obtained. In this image, phase differences are only due to atmospheric delay and
perhaps long wavelength trends due to orbit inaccuracies. The latter can be approx-
imated using the variance in the orbital state vectors and tie-points at the surface.
In this procedure, all long wavelength (> 200 km) atmospheric information is ef-
fectively eliminated. The residual differential interferogram can be regarded as an
“atmosphere-only” image.
Phase differences in the differential interferogram need to be interpreted with some
precaution. First, it is important to realize that the interferogram only contains
relative phase information. There is no absolute calibration point. As a result, the
value of one single pixel is useless. Second, the variation of the phase is due to the
spatial variability during two acquisition times with a different state of the atmo-
sphere. This ambiguity is a limiting factor for data interpretation if only two SAR
images are used, as in this study. Using a number of SAR images, interferograms
238 Chapter 6: Atmospheric monitoring

can be made between different combinations, which enables one to overcome this
ambiguity. For this study, the sign of the phase values is used to attribute the phase
effect to the SAR image of 23 April or 24 April. This approach can be used for
strong disturbances—for small phase variation it is not possible to attribute it to
one specific acquisition using only two images.
Using eqs. (6.1.3) and (6.1.4) slant and zenith delay differences are derived from the
phase differences. Assuming smooth lateral pressure and temperature variation over
scales less than 50 km, the observed delay variation is mainly due to variations in
the wet part of the refractivity. Using the method proposed by Bevis et al. (1992)
precipitable water variations, labeled differential precipitable water are
derived from the delay interferogram and surface temperature measurements. After
 6.7 Spatial moisture distribution during CLARA 96 239

geocoding the map from radar coordinates to the WGS84 ellipsoid, the result
is shown in fig. 6.18.
To validate the meteorological interpretation of the map, additional datasets
from the CLARA experiment are used. GPS zenith delay observations, radiosonde
profiles, lidar, infrared and microwave radiometer data were available at the Delft
station, indicated by the black dot in fig. 6.18. Additional Meteosat and NOAA-
AVHRR data were used for comparison as well.
For the analysis of the spatial variability of the delay or water vapor signal, the
two-dimensional structure function, defined as

is used. Here, is the zenith delay signal and and are the distances between
two arbitrary points in km. The structure function is the expectation value of the
squared difference between two points at a certain distance R and azimuth in the
image. It exists for all random functions with stationary increments. In section 4.1.4,
the structure function was discussed in more detail.

6.7.2 Results

Here, a first interpretation of the interferogram is made, followed by the general

meteorological interpretation of both days using the existing CLARA ground truth
data. Then, a comparison with the GPS data and a statistical evaluation is per-
formed.
Interferogram interpretation
Figure 6.18 shows the differential PWV map for 23 and 24 April 1996. The colorbar
is linear between –1.5 and 1.5 mm in order to use the full color resolution and
suppress the effect of a few localized points with very high values. The yellow-
red range of the colorbar corresponds with delay signal on 24 April, blue shades
correspond with 23 April. The variation can be interpreted as due to water vapor
and clearly indicates bandedness along the wind direction on 24 April: SW–WSW.
In the lower-left part of the map, it can be observed that moisture converges
into linear bands, which expand to the large anomaly at 51.4°N, 4.3°E. The value and
shape of this anomaly indicates a deeper column of moist air, which is expected for
cumulus–cumulonimbus clouds and for regions penetrated by precipitation. In those
sub-cloud regions, the relatively dry air will become saturated by evaporating liquid
particles. Although it is in general not possible to determine whether precipitation
occurs from the interferometric data only, localized showers often give a strong signal.
The shape and magnitude of the anomalies suggest very localized concentrations of
water vapor, as expected for convective weather with developing cumulus clouds.
The blue regions, e.g., at 51.8°N, 4.0°E and 51.5°N, 4.7°E, corresponding to 23 April
also indicate strong concentrations of water vapor. The existence of clouds in the
area is very likely, although probably with less variation than at 24 April.
In fig. 6.19A, the weather radar reflectivity of both days, 10:30 UTC, is shown,
converted to rain rate in mm/h. To facilitate comparison with the interferogram,
240 Chapter 6: Atmospheric monitoring

precipitation during 23 April is plotted using negative (blue) colors, and precipitation
during 24 April using positive (yellow) colors. For the grid points where precipitation
occurred during both days, the average value is shown. The correlation between the
precipitation regions and large values is remarkable, although the absence of
precipitation at 51.8°N, 4.0°E supports the hypothesis that the signal can
be caused by towering cumulus alone.
General interpretation based on additional data
According to the weather maps of 23 April, a cold front has just passed over the
interferogram area during the SAR acquisition. There is no obvious frontal passage
signature in neither the satellite imagery nor the Delft soundings nor the Delft
surface measurements of temperature, humidity or pressure, thus the cold front
was quite weak. Behind it, there are isolated showers in a neutral atmosphere,
as derived from weather radar and soundings. Cloud base is 2 km, as determined
from the IR radiometer cloud base temperature of 5°C, which corresponds to 2
km on the sounding. This is confirmed with the ESTEC lidar. The radiosonde
data reveal strong winds, 10–15 m/s, from SSW. Cloud types are a combination
of stratocumulus and cumulus. Near the time of the frontal passage at Delft, there
were more clouds and they were more continuous (determined from liquid retrieval
of microwave radiometer; also IR radiometer shows continuous cloud cover), and
there was a maximum in water vapor. Then cloud cover diminished as the water
vapor amounts decreased.
 6.7 Spatial moisture distribution during CLARA 96 241

During the acquisition of the second SAR image, at 24 April, there were more isolated
showers, but not as many as during the first day. The radiosonde observations
indicate a more stable atmosphere with colder, drier air at the surface. The cloud
base is 2500 m (determined from the ESTEC lidar), which is consistent with the
IR radiometer cloud base temperature of –10° C and sounding temperature at this
height of about –6°C. The cloud cover is thin and broken, and consists of mostly
cumulus, as indicated by the IR radiometer, the ragged look of the liquid water
retrieval from the microwave radiometer, and by the order of magnitude less in
liquid water content from the microwave radiometer on 24 April compared with
23 April. Winds are weaker during this day, 8–10 m/s, from WSW. Cloud streets,
mainly developing cumulus, are visible over the interferogram area in the Meteosat
double visual image at 10:30 UTC.
Comparison between SAR and GPS data
GPS observations were performed at station Delft during the two days of the interfer-
ogram. In fig. 6.19B, the derived zenith wet delay is shown using a 6-min sampling
interval. Hydrostatic delay components are removed using surface pressure data.
Two pairs of vertical lines indicate an interval of 2.45 hours around the SAR acqui-
sitions. This interval was determined from an average wind speed of 10 m/s and the
interferogram size of 100 km.
From the signal variation within these two intervals, it is expected that the signal
variation in the interferogram is less than 15 mm. However, the zenith wet delay
variation in the interferogram has a range of 54 mm. Table 6.3 lists some statistical
values for SAR and GPS.
From fig. 6.1.9B and table 6.3 it is obvious that a comparison between 24 hours of
GPS data and the interferometric data is not possible, as the daily variation in wet
delay is much stronger than the variation over a short interval. This behavior is also
expected from power-law considerations. On the other hand, only evaluating the 2.45
hours observations centered around the SAR acquisition time is not representative
as well, as it seems to underestimate the amount of variation.
The rms of the zenith wet delay variation in one SAR image is assumed to be equal
to the variation in the interferogram divided by or 3.7 mm.
242 Chapter 6: Atmospheric monitoring

Regarding the spatial resolution of the GPS zenith delay observations, two remarks
need to be made. First, for a cloud base at 2 km and a 20° elevation cut off, the di-
ameter of the GPS cone at that altitude is 11 km. If an array of GPS receivers would
be installed with a 11 km posting, independent observations would be guaranteed at
that altitude. From Shannons sampling theorem it follows that water vapor signal
with a wavelength of 22 km or less will be aliased into the longer wavelengths. For
the SAR-derived spatial delay field, using a spatial averaging to 160 m, the obser-
vations can be considered independent. On the other hand, for wavelengths larger
than, say 50 km, gradient errors due to satellite orbit inaccuracies and hydrostatic
effects limit the unambiguous interpretation of the signal. Such considerations can
be important when an array of GPS receivers is available.
Secondly, when only one GPS receiver is available, the temporal behavior of the sig-
nal might give a first indication of the delay variability. To perform the conversion
from temporal to spatial information, Taylors hypothesis of a “frozen” boundary
layer is assumed in which the anomalies do not develop but are only transported in
their original shape by the wind (Taylor, 1938). With this rather stringent assump-
tion, the wind speed is needed for the conversion. The observed wind speed during
23 April is approximately 10 m/s. With the 11 km diameter of the GPS observa-
tion cone at 2 km altitude, the sampling interval should be 18 minutes to interpret
the observations as independent. The 6–10 min sample interval used during the
CLARA experiment is therefore clearly too short to interpret the short-wavelength
signal. For a comparison with the SAR observations, the sampling interval of 20 min
should be used, which results in a shortest wavelength of 40 min, or 24 km.

Scale dependent statistics

For the description of boundary-layer (200 m–2 km) turbulence characteristics as well
as analyzing the geodetic implications of positioning using space-geodetic methods,
the spatial behavior of the delay variation is very important. For the relative radar
interferogram the first moment or expectation value of the delay is zero. In fig. 6.20,
the histogram of the delay data from the interferogram is displayed. The second
moment or dispersion of single observations can be derived but is less interesting as
the dispersion of increments. The latter can be described by analyzing the power or
the magnitude spectrum or by the structure function and reveals information on the
decrease in power for smaller spatial increments. Figure 6.21A is the rotationally
averaged amplitude spectrum of a continuous part of the interferogram. It is derived
by computing the 2D FFT and averaging all values with the same radial distance
to the origin. The diagonal lines indicate the typical –5/3 power law decay for
higher frequencies. Note that the –5/3 slope observed in a rotationally averaged
spectrum corresponds with a –8/3 slope in a one-dimensional spectrum, consistent
with our results in section 4.7 (Turcotte, 1997). For wavelengths between 0.5 and
2.5 km the variation clearly follows this decay. A small deviation at 2.5 km (0.4
cycles/km) could indicate the effective height of the boundary layer in which 3D
turbulence characteristics play a dominant role. For longer wavelengths, the decay
is less strong (approaching –2/3 for the rotationally averaged or –5/3 for a one-
dimensional signal) as expected for effective 2D turbulent characteristics.
 6.7 Spatial moisture distribution during CLARA 96 243

The two-dimensional structure function reveals information on the corre-

lation of single pixels with other pixels in the image. In fig. 6.21B, the square root of
the structure function is shown. The non-isotropy of the field can be clearly derived,
in this case mainly caused by the two strong delay signatures in separate acquisi-
tions. As the image is derived from averaging all combinations of pixels, there are
less realizations available for the edges and corners of the image. Therefore, areas
outside 70 km from the origin need to be interpreted with care. For geodetic ap-
plications, it is important to increase the variance of phase difference measurements
with increasing distance.

6.7.3 Discussion

Radar interferometric data can be used to obtain a relatively accurate and high-
resolution interpretation of the state of the boundary layer. Of all the observational
datasets available for the CLARA project, none of them give a two-dimensional
spatial overview. Weather radar gives a spatial view only if precipitation occurs,
while Meteosat and NOAA-AVHRR lack resolution to observe the fine details in
water vapor. All other instrumentation is situated at a single point and gives tem-
poral or vertical information. Therefore, the radar interferograms add considerable
complementary value in terms of high resolution and quantitative information.
A disadvantage of many new sensors (e.g., water vapor DIALs) is that they do
not work in cloud. Therefore, radar interferometry has an advantage with respect
to that as well. In the scientific community there is strong interest in water vapor
measurements within clouds (Weckwerth et al., 1999). There may be supersaturation
or subsaturation within clouds, which affect droplet size distribution and particle
type and therefore the radiation. With existing instrumentation, this is very difficult
to determine.
Nevertheless, major drawbacks are currently in the unambiguous interpretation of
244 Chapter 6: Atmospheric monitoring

two superposed atmospheric situations. With only two SAR images to form one
interferogram, this will be a limiting factor. Suitable SAR platforms with more
frequent revisit times are expected to solve this problem. Moreover, design consid-
erations might result in, e.g., larger swath widths or reduced resolution to speed up
data processing.

6.8 Conclusions

Satellite radar interferometry can be applied to study vertically integrated atmo-

spheric refractivity variations with a spatial resolution of 20 m and an accuracy
of ~2 mm, irrespective of cloud cover or solar illumination, over most land areas.
Therefore. Interferometric Radar Meteorology (IRM) is a new technique for me-
teorological applications and atmospheric studies. The data are derived from the
difference between the radar signal delay variations within the imaged area during
two acquisitions with a temporal separation of one or more days. Hence, they reflect
the superposition of the refractivity distribution during these two acquisitions. On
short spatial scales integrated refractivity variations appear to be dominantly caused
by spatial heterogeneities in the water vapor distribution, only slightly modulated
by influences of liquid water and temperature. Hydrostatic delay gradients can be
commonly ignored, as well as ionospheric influences for C-band radar.
The configuration of contemporary SAR satellites is currently far from optimal for
 6.8 Conclusions 245

operational meteorology. This is mainly due to the generally infrequent repeat-

acquisition interval. A related problem is the superposition of two atmospheric
states. It has been shown, however, that the interferometric combination of a series
of SAR acquisitions can be used to identify the atmospheric influence per single ac-
quisition (Ferretti et al., 2000). Power consumption, high data rates, and significant
signal processing effort form another challenge for real-time operational applications.
Nonetheless, IRM can be readily used for studying atmospheric dynamics such as
mesoscale shallow convection (MSC), and provides unsurpassed insights, particu-
larly in mapping the small-scale water vapor distribution. The main advantage of
the technique is the combination of its imaging possibilities with quantitative de-
lay estimates. ESA’s tandem mission (August 1995–April 1996) acquired more than
110.000 tandem interferograms covering nearly the total global land surface (Duchos-
sois et al., 1996). Since surface deformation is practically excluded in these data,
and topography can be considered known for many areas (especially after the data
release from the SRTM mission), almost all of these interferograms can be used for
atmospheric studies.
This chapter shows a number of case studies that validate the radar interferometric
results based on the comparison with weather radar, spaceborne radiometer, GPS,
radiosonde, and synoptic (acquired at the same time around the world) ground-based
observations. For a specific situation, a cross-validation combining GPS, radiome-
ter, and InSAR proves possible, resulting in quantitative measures of water vapor
content. It proves possible to study causal connections between mesoscale wind pat-
terns and moisture distribution in the lower boundary layer by combining the radar
backscatter intensity over water and the radar signal delay over land areas. Analyses
are reported on the relation between wind and moisture associated with boundary
layer rolls and with a cold-frontal rain band.
There are a number of limitations for the application of IRM. First, the technique is
only applicable over land or ice areas. Second, whereas the time interval between the
acquisitions should be as short as possible to reduce possible temporal decorrelation,
it has to be long enough to allow for the necessary atmospheric decorrelation. In
other words, there needs to be some change in the atmospheric refractivity distribu-
tion. The latter condition can be somewhat relaxed if acquisitions are available very
frequently. Third, topography needs to be known as a DEM with a vertical accuracy
of ~10 m and a spatial resolution as fine as the SAR resolution. Using the results of
the SRTM mission, this is currently feasible for latitudes between 58°S and 60°N.
If the area is characterized by large elevation differences, atmospheric signal due to
vertical stratification needs to be accounted for. Fourth, satellite orbits need to be
of sufficient quality to remove residual phase gradients as good as possible, and to
prevent DEM scaling errors. Very reliable orbit control allows for the retrieval of
hydrostatic gradients. An ideal future system would benefit from a wide swath, e.g.
by applying a ScanSAR mode, and could operate with a reduced spatial resolution.
Such a system would have unique complementary characteristics with conventional
meteorological techniques.
 Chapter 7
Conclusions and recommendations

The general problem statement of this study was formulated in chapter 1 as:
"How can the interpretation and analysis of repeat-pass spaceborne radar in-
terferometric data be improved in a systematic way via a model-based quantifi-
cation of the error sources?” Here we present the main contributions of this
research and recommendations for further studies and practical application of
radar interferometry as a geodetic technique.
key words: Conclusions, Quality assessment, Error analysis and propagation,
Atmospheric signal

This study discussed the technique of repeat-pass spaceborne radar interferometry

from a geodetic point of view. The main conclusion of this research is that the
interpretation and analysis of radar interferometric data can be improved by formu-
lating the problem using a standard Gauss-Markoff model and by including the error
sources as variances and covariances in the stochastic part of this model. Especially
covariance modeling is an important extension of current quality assessment strate-
gies. Second, this study presented a novel technique for meteorology, referred to as
Interferometric Radar Meteorology (IRM), which provides fine-resolution maps of
water vapor distribution.
The conclusions in relation to the four specific research questions listed in chapter 1
are discussed below, followed by the recommendations for further research.

7.1 Contributions of this research

7.1.1 Model formulation

Specific research question 1. “How can we parameterize the formation

and the quality of a radar interferogram and its derived interferometric
products?”

The general problem statement in chapter 1 was motivated by the need for a geodetic
formulation of the observation equations and of the variance-covariance matrix of the
observations. A Gauss-Markoff model (see chapter 3) can be used as a framework to
combine both the functional relations between observations and parameters and the
248 Chapter 7: Conclusions and recommendations

stochastic characteristics of the observations. This formulation allows for a direct

evaluation of the degree of redundancy, quantitative statements on precision and
reliability, and standard methodology for data adjustment and filtering.
In the generic functional part of the model, every interferometric phase observation
can be related to at least five parameters: topographic height, surface deformation,
atmospheric signal during the two acquisitions, and the integer phase ambiguity,
hence observations are related to parameters. Based on this generic (underde-
termined) model, the specific model for topographic mapping, deformation mapping,
or atmospheric mapping can be derived, taking into account the geometric or tem-
poral characteristics of the acquisitions as well as the availability of supportive data
such as a digital elevation model or meteorological information.
By considering atmospheric signal as additive noise on the phase observations, it
is possible to exclude it from the functional model, and include it in the stochastic
model. As such, atmospheric signal results in a dramatic increase in the variance
of the observations combined with the introduction of covariances as a function of
distance.

7.1.2 Error sources and propagation

Specific research question 2. “Which error sources limit the quality of

the interferogram and how does their influence propagate in the interfer-
ometric processing?”

Error sources that limit the quality of the interferogram have been discussed in chap-
ter 4. Single-point observation statistics show that the probability density function
of the interferometric phase can be expressed as a function of the coherence and the
number of effective looks. For distributed scattering mechanisms, coherence can be
estimated directly from the data, although this hampers the identification of single
coherent point scatterers. Moreover, it is shown that the coherence estimators are
biased and have an increased variance for low coherence values and a low number
of effective looks. Hence, there is a strict trade-off between coherence estimator
accuracy, i.e. phase variance estimation, and spatial resolution. In fact, point scat-
terers may have a phase variance which is sufficiently low to allow for reliable phase
estimates.
A system theoretical estimate of coherence can be obtained for geometry-induced ef-
fects such as geometric (baseline) decorrelation and Doppler centroid decorrelation.
Coherence effects induced by interferometric processing procedures, such as coregis-
tration, interpolation, complex multiplication, and filtering can be treated the same
way. The effects of thermal noise can be derived from the design parameters of the
SAR system and estimates for the normalized radar cross section.
Apart from a highly unlikely micro-scale modeling of the scattering characteristics
for a specific area, it is impossible to predict the amount of temporal decorrelation
for each resolution cell quantitatively. Many indications, such as land use, vegeta-
tion, anthropogenic activity, climatic conditions (such as snow), and possible surface
deformation can give an indication of the amount of decorrelation to be expected, in
 7.1 Contributions of this research 249

combination with the radar wavelength. Nevertheless, these parameters are usually
not quantitative and depend on operator experience and a priori information. The
propagation of the various sources of decorrelation error to the phase variance is
summarized in fig. 4.39, p. 155.
An important characteristic of the interferometric phase is the of its
PDF. If we intend to refrain from non-verifiable assumptions on the maximum
phase gradient, the location of the phase PDF is unknown. By introducing pseudo-
observations, the range of the PDF can be limited, and by using a sufficient number
of aligned SAR images with varying spatial and temporal baselines, an optimal
solution can be obtained. These topics are outside the scope of this study, but will
be of major importance for future research. For many interferometric data analyzed
here, heuristic assumptions on maximum phase gradient were used, allowing for
“traditional” phase unwrapping algorithms to be applied.
Correlation between the phase observations of resolution cells in the image is intro-
duce by orbit errors and variable atmospheric delay. The effect of orbit (baseline)
errors is shown to have long-wavelength characteristics for the reference phase and
a scaling error when deriving topographic height differences. It is shown that a cor-
rection of the baseline by counting the residual fringes may result in an increased
error in the perpendicular baseline. The amount of residual fringes based on the
a priori estimation of the radial and across-track orbit errors are approximated.
Errors in the interferometric phase due to the atmospheric refractivity distribution
can be distinguished into two contributions: turbulent mixing and vertical strati-
fication. Turbulent mixing induced refractivity distribution results in a power-law
(fractal) distribution of the spatial delay variability. Such spatial variability can
be conveniently described using structure functions or covariance functions. An
atmospheric delay model is introduced based on Kolmogorov turbulence theory. Ex-
perimental data over many interferograms support this model.
Differences in vertical stratification between the two SAR acquisitions significantly
influences the interferometric phase. An empirical relation between the size of the
phase error due to vertical stratification has been derived for different height in-
tervals and temporal baselines, using a set of radiosonde observations. Correction
of the error due to vertical stratification is only possible using vertical profile mea-
surements. Surface observations combined with a tropospheric model are in general
unreliable, whereas integrated refractivity observations as obtained using GPS are
insufficient for correction.
Propagation of errors to the final interferometric products, i.e., a deformation or
elevation map, is relatively straightforward and is discussed in section 4.9. An im-
portant contribution of this research, with respect to previous studies on InSAR
accuracy, is the parameterization of the errors in terms of both covariances and vari-
ances. For example, the height difference between two pixels in an elevation map is
generally more accurate if the pixels are close together. This is valuable informa-
tion for the end user of the map, which now becomes available as a complementary
product.
250 Chapter 7: Conclusions and recommendations

7.1.3 Modeling of atmospheric error

Specific research question 3. “To what extent does the atmospheric re-
fractivity distribution affect the interferograms. Is it possible to deter-
mine this influence using additional measurements, and how should it be
modeled stochastically?”

The influence of the atmospheric refractivity distribution can significantly affect

repeat-pass radar interferograms. A study performed within the framework of this
research shows that each interferogram of a series of 26 tandem acquisitions over the
Netherlands is affected by atmospheric signal (Hanssen, 1998). The observed spatial
scales reach from hundreds of meters to 100–200 km. For scales reaching the full
image size, it is, however, difficult to discern them from errors in the satellite orbits.
Although rms-measures of the magnitude of the effects do not describe the spatial
roughness of the signal and are a function of the spatial extent of the evaluation area,
they yield a comprehensible approximation of the atmospheric signal. Observed
rms-values, taken over an area of approximately 40 × 40 km, reach from 2–16 mm.
Assuming a Gaussian distribution, this implies that 95% of the observed phase
values varies over a minimal range of 8 mm up to a maximal range of 64 mm.
Extreme ranges of more than 120 mm delay variation are observed in the case of
thunderstorms during the two SAR acquisitions.
It is shown that (section 4.7) although the state of the atmospheric refractivity dis-
tribution can be very different depending on the weather situation, the shape of
the power spectrum shows three distinct scaling regimes, which are similar indepen-
dent of the state of the atmosphere. Nevertheless, the absolute values of the power
spectra differ two orders of magnitude. A model for atmospheric delay variation is
presented, which incorporates these scaling characteristics, and which can be ini-
tialized for various weather or climate situations. Using the model, it is possible to
quantitatively describe the covariance function of an interferogram using only one
parameter. This parameter can be estimated from a small patch of the interfero-
gram, with either even terrain or without deformation—the bootstrap method. This
first order estimate may then be used to describe covariances in the whole interfero-
gram. An alternative approach is the estimation of the initialization parameter using
GPS time series—GPS initialization. In this way, a single GPS receiver in the inter-
ferogram area would be sufficient to quantify the covariances in the interferometric
data.

7.1.4 Interferometric radar meteorology

Specific research question 4. “Is it possible to analyze the atmospheric

error signal as a meteorological source of information?”

Interferometric Radar Meteorology (IRM) is a new technique for meteorology and

atmospheric studies in general. It can be applied to study vertically integrated at-
mospheric refractivity variations with a spatial resolution of 20 m and an accuracy
of ~2 mm, irrespective of cloud cover or solar illumination, over most land and
 7.2 Recommendations 251

ice areas. On short spatial scales integrated refractivity variations are dominantly
caused by spatial heterogeneities in the water vapor distribution. IRM can be read-
ily used for studying atmospheric dynamics such as mesoscale shallow convection,
by mapping the small-scale water vapor distribution. The main advantage of the
technique is the combination of its imaging possibilities with quantitative delay es-
timates. Readily attainable integrated precipitable water amounts over large areas
may make pinpoint weather forecasting a possibility. The delay maps could eventu-
ally be used by bench forecasters or serve as an additional constraint in variational
data assimilation models.
A number of case studies validate the radar interferometric results based on the
comparison with various meteorological data and GPS data. The combination of the
instantaneous observations of radar backscatter intensity over water and radar signal
delay over land areas enables the study of causal connections between mesoscale
wind patterns and moisture distribution in the lower boundary layer. Analyses are
reported on the relation between wind and moisture associated with boundary layer
rolls and with a cold-frontal rain band.
Contemporary satellites, however, have an orbital repeat period that is far from
ideal for operational applications, resulting in loss of coherent phase information
over many areas. Ideally, SAR data should be acquired with multiple short tem-
poral spacings in order to maintain high coherence and permit a large number of
interferogram pairs to be generated. Multiple observations give not only better
precision in time, but keep coherence and hence data quality high. Proper use of
airborne and spaceborne SAR systems with short repeat periods could lead to much
greater accuracy in meteorological understanding and forecasting.

7.2 Recommendations

The mathematical model introduced in this study is only a first step towards a unified
approach for the analysis of radar interferometric data. A first implementation of
this theory in software algorithms is necessary to validate the concepts.
The implementation of the theory discussed in this study in operational algorithms
can be performed in two stages. In the quality assessment stage, the accuracy of (i) a
single pixel parameter and (ii) the difference between two pixels can be determined,
assuming distributed scattering. For example, if the final product is an elevation
map it will be possible to choose an arbitrary pixel and obtain both the height
and variance of the height for that pixel. More important, it will be possible to
choose two pixels in the image and obtain the height difference and the variance
of the difference, using The same reasoning holds for
deformation mapping or atmospheric mapping. The quality assessment stage is
relatively easy to implement by combining the coherence image, the interferometric
product derived from the interferogram, the atmospheric initialization parameter
and the a priori estimates for the radial and across-track orbit errors.
The second stage involves the implementation of the variance-covariance matrix.
Using a full-scene single-look interferogram, this will be a matrix with millions of
elements. Matrix operations such as inversions will be vary laborious, if possible at
252 Chapter 7: Conclusions and recommendations

all. Nevertheless, the well-organized structure of the matrix (BTTB) suggests that
standard inversion procedures might not be the most economic. For example, it
could be investigated whether it is possible to derive the weight matrix directly
based on theoretical grounds. These issues need to be considered, combined with
problems on numerical stability and parallel processing.
The PDF of the interferometric phase has a As data adjustment using
a Gauss-Markoff model is conventionally limited to observations with an unambigu-
ous PDF. this poses a limitation in using the model for wrapped data. The use of
more than two interferograms for a specific goal (e.g., topography or deformation)
results in a different sensitivity for height or deformation, depending on the spatial
and temporal baseline, respectively. Integration of this information in the model
will result in a more generic approach in the data adjustment.
Although the different sensitivities for topography and deformation result in different
scaling of the PDF, it theoretically still extends to infinity. Therefore, pseudo-
observations could be introduced, which decrease the likelihood of solutions outside
the physical reality. It needs to be investigated how these pseudo-observations need
to be structured and incorporated into the model.
The use of many, or all available, SAR acquisitions for a certain area for analyzing
a specific problem will become more important. The combination of time-series of
interferograms (stacking) with the spatial deformation pattern within an interfero-
gram has already lead to new applications of the technique, previously unattainable.
In this light, the use of the permanent scatterers technique is a major contribution,
which deserves more attention (Ferretti et al., 2000).
Finally, regarding deformation measurements there appears to be a gap between
the derived slant-range deformation pattern and the determination of geophysical
parameters. Often, forward models are applied to minimize the difference between
the observed and the modeled deformation pattern. It needs to be investigated
whether the proposed geodetic model discussed in this study can be converted to
directly derive the geophysical parameters of interest. The advantage of such an
approach is that accuracy and reliability can be directly derived from observations
and model, and that model adjustment is possible based on the characteristics of
the observations.
Addenda
 Appendix A
Comparison neutral delay GPS and InSAR

Neutral (tropospheric) delay can be an important source of error when using

radar interferometry for deformation or topographic studies. Since the signals
of the Global Positioning System are affected by the troposphere in a similar
way as the radar signals, it is investigated whether GPS data can be used for the
correction or quantification of the tropospheric error in radar interferometric
data. Two experiments discuss the use of a spatially distributed set of receivers
and of one stationary receiver, using time series analysis.
key words: GPS, Troposphere, Signal delay, GPS network, GPS time series

Atmospheric delay in radio signal propagation is known to be a major source of error

in high-precision positioning applications such GPS and InSAR (see, e.g. Bevis et al.
(1992) for GPS and section 4.7 and chapter 6 for InSAR measurements). In GPS
data processing, the dispersive ionospheric part of the delay is largely eliminated by
using a linear combination of the two GPS frequencies. The non-dispersive delay
in the neutral part of the atmosphere cannot be eliminated this way, and is often
estimated as an extra unknown in the GPS data processing (Brunner and Welsch,
1993). In this procedure it is often assumed that there are no horizontal variations
and that the observations can be related to zenith delay using a mapping function
(Herring, 1992; Niell, 1996). If the tropospheric delay is the main parameter of
interest, it is often decomposed in a hydrostatic (“dry”) component, which can be
estimated with a 1 mm accuracy using surface pressure measurements, and a “wet”
component which is highly variable, both spatially as well as temporally (Davis
et al., 1985). Nowadays, the wet delay can be derived from the GPS observations
at a fixed station with an accuracy of about 6–10 mm in the zenith direction (Bevis
et al., 1992; van der Hoeven et al., 1998).
Since the troposphere is not dispersive for frequencies less than 30 GHz, the GPS
signal delays are affected the same way as, e.g., X-band, C-band, or L-band SAR
observations. Therefore, if GPS stations are available in the interferogram area,
it might be possible to use the GPS-derived zenith delays as a priori information
in the interpretation of the interferogram. There are two possibilities to combine
GPS and InSAR observations. In the first, deterministic, approach the relative delay
differences between the stations at the time of the image acquisitions are determined,
interpolated, and subtracted from the interferogram data. The second, stochastic,
256 Appendix A: Comparison neutral delay GPS and InSAR

approach does not aim at full correction, but uses time series observed at a stationary
receiver to estimate the amount and variability of delay variation. This way, a
quantitative stochastic measure of the delay behavior can be obtained. Experiments
which investigate both possibilities are discussed in the sequel. Section A.1 explores
the possibilities of a spatial network of receivers, whereas section A.2 covers the
single stationary receiver.

A.1 Spatial networks

The experimental data for this study were collected during the Groningen Interfer-
ometric SAR Experiment (GISARE), performed during 14–18 March 1996 in the
northeast of the Netherlands. In the framework of this experiment tandem SAR
data were acquired during five months by the satellites ERS-1 and ERS-2. During
two of these acquisitions, simultaneous GPS measurements were performed using a
temporary network of 8 GPS receivers. By comparing the GPS zenith signal delays
with the interferometric data, an analysis can be performed of the correspondence
and differences, regarding spatial and temporal resolution, accuracy, complementar-
ity and feasibility (Hanssen, 1996: Stolk, 1997; Stolk, Hanssen, van der Marel and
Ambrosius, 1997).
The difference in the interferometric phase between two points m and k, in terms of
tropospheric delay at acquisition can be written as
 A.1 Spatial networks 257

where is the SAR wavelength. Using a simple mapping function,

with local incidence angle the tropospheric delays can be converted
to zenith delays:

The interferometric phase values represent the temporal difference in tropospheric

delay, but only in a relative sense. By comparing different locations and in the
interferogram, the spatial distribution of these differences can be studied.
The relative tropospheric delays can be derived from the GPS phase observations as
well. Clock errors are eliminated using double-difference observations. Ionospheric
delays are eliminated by using the ionospheric-free linear combination (L3). The
effective wavelength of L3 is denoted by Assuming known satellite ephemerides
and solving for the integer phase ambiguities, the double-difference residuals can be
obtained:

where is the double-difference tropospheric delay between stations and

and satellites and at Multipath is denoted by and measurement
noise by . Separating the double differences to the two stations and applying
a mapping function to derive zenith delays results in

with

Mapping function M is a function of the elevation angle to the satellites. In

eq. (A.1.4) the tropospheric delay occurs in two terms: (i) the relative spatial tro-
pospheric zenith delay difference between station and and (ii) in the absolute
value of the tropospheric zenith delay for station Both terms can in principle
be derived from the double-difference residuals of a measurement epoch. However,
the coefficient of the absolute term is dependent of the length of the baseline
between and relatively short baselines imply nearly identical elevation angles
for both stations and both satellites. Therefore, for these baselines, and no
information on the absolute delay can be obtained. A similar problem can occur for
coefficient in case both satellites are observed at the same elevation angle. In these
cases (which are only a temporary problem) the relative tropospheric delay difference
cannot be determined. In general, however, it is possible to derive estimates for rela-
tive tropospheric delay differences from double-difference, ionosphere-free residuals,
independent of the length of the baseline.
258 Appendix A: Comparison neutral delay GPS and InSAR

Comparing the GPS relative tropospheric delay differences at the SAR acquisition
times and and assuming that multipath effects will cancel in the differencing
of a 24 hr interval we find

Comparing the right-hand side of eq. (A.1.5) with the right-hand side of eq. (A.1.2)
and assuming zero-mean Gaussian noise we find

hence, both the GPS and SAR observations can be compared.

A.1.1 Experiment set-up

During the GISARE-GPS experiment, between 14 and 18 March 1996, simultaneous

GPS observations where performed on eight NAM locations in the northeast part
of the Netherlands. Using a survey epoch of 5 days, it is possible to determine
the position of the receivers with an accuracy higher than 1.2 mm. Eight GPS
receivers (4 A shtech MDL Z-12, 4 Turbo-Rogue’s SNR-12RM) with Dome-Margolin
antennas were used simultaneously during this period. At 16 and 17 March the
area was imaged by ERS-1 and ERS-2, respectively, see figs. A.1 and A.3. Since
the perpendicular baseline was 26 m, the sensitivity for topographic height is very
small, see 2.4.1, excluding this influence on the interferometric phase. Since the data
were acquired within 24 hours surface deformation can be excluded as well, making
atmospheric signal the most prominent phase signal.
 A.1 Spatial networks 259

A.1.2 Methodology

To retrieve the double-difference ionosphere-free residuals from the GPS phase ob-
servations, it is necessary to determine the positions of the satellites and receivers,
and to solve for the integer ambiguities. Satellite positions where made available by
the International GPS Service for Geodynamics (IGS). Using the Bernese GPS soft-
ware all receiver stations are tied to a network of independent baselines, to solve for
the station coordinates and the integer ambiguities (Rothacher and Mervart, 1996).
For every baseline, the double-difference ionosphere-free residuals are determined.
Since the baselines are relatively short, it is not possible to determine the absolute
tropospheric delay. It is, however, possible to estimate the relative delay differences.
A relative tropospheric delay is estimated every 30 seconds, using a cut-off elevation
angle of 20°. For every baseline, this results in a time series spanning one day. To
eliminate noise in these time series, the data are low-pass filtered. From these fil-
tered data, the value of the relative tropospheric delay is recorded at the time of the
SAR acquisitions.

The interferometric data are multilooked by 2 × 10 pixels, resulting in a spatial

posting of approximately 40 × 40 m. After phase unwrapping, the phase difference
between adjacent pixels can be interpreted as a spatial difference in tropospheric
delay during the two acquisitions. Since orbit errors can result in a residual phase
ramp this is accounted for in the comparison with the GPS data.

The part of the atmosphere that contributes to the observed delays is different for
GPS and InSAR. The atmospheric contribution for a GPS receiver is a reversed
cone, with the receiver in the focal point and the imaginary rays to several GPS
satellites located within or at the edges of the cone, see fig. A.2. The aperture angle
of the cone is defined by the cut-off elevation angle, below which observations are
discarded. All phase observations between receiver and the satellites, averaged over
a certain time interval, contribute to the derived zenith delay observation. For SAR
interferometry, the delay observation at a single pixel can be regarded as one single
ray, ranging from that pixel to the satellite, at the time of the image acquisitions.
These observations are spatially relative, and are formed by differencing the delays
at two, nearly uncorrelated, acquisition times. To ensure analysis of comparable
quantities, both data sources need to be tuned.

Regarding both sources of data, it is assumed that especially the local variations in
the wet delay, dominantly water vapor, contribute to the relative differences. The
analysis of radiosonde data acquired at noon shows that more than half of the wet
delay originates from the lower 1.4 km of the troposphere. The intersection of the
cone and the 1.4 km level forms a circle with a diameter of 7.6 km, and it is assumed
that atmosphere within that circle is the main contribution to the GPS wet zenith
delay observation. To tune the high-resolution interferometric data to the same area,
the interferogram is averaged over all data within that circle, considering a lateral
shift of the circle due to the SAR incidence angle. This geometry is visualized in
fig. A.2. As a result, the SAR data regard a larger fraction of the atmosphere below
1.4 km, whereas the GPS data consider a larger atmospheric volume above this level.
260 Appendix A: Comparison neutral delay GPS and InSAR

A.1.3 Results

Figure A.3 shows the full-scene interferogram of the SAR acquisitions of 16 and
17 March 1996. The area covered by the eight GPS receivers covers only the
southeastern quadrant of this image. Atmospheric signal is limited to a weak long-
wavelength anomaly in the southeast (the dark part) and short-wavelength waves
over the larger part of the image. The latter are probably caused by gravity waves—
waves at the boundary between two masses of air with different densities—or due to
wind shear. Local variations of less than 1 cm are observed, while the wavelength
of the gravity waves is ~250–300 m.
The averaged areas around each GPS receiver, used for the quantitative comparison
with GPS are indicated in fig. A.3, and consist of ~25000 pixels. The difference
between two of these circles can be compared with the GPS observations related to
that specific baseline.
Additional meteorological information is obtained from Meteosat and NOAA-AVHRR
(Advanced Very High Resolution Radiometer) satellite imagery, synoptic surface ob-
servations, radiosonde launches, and weather radar, see Hanssen (1998). Both days
had full (8/8) cloud cover during the SAR acquisitions. Stratus and stratocumulus
clouds were observed at 300–400 m, while during the second acquisition also higher
cloud cover was observed. The AVHRR images show varying wave patterns over
 A.1 Spatial networks 261

the test area. The wind direction varied from east at the surface to south aloft.
Temperatures during the SAR acquisitions were a few degrees above the freezing
level.
Of the eight installed GPS receivers, one failed after some hours of operation. Of
the seven remaining stations, six independent baselines are defined. Relative tropo-
spheric delay differences can be derived for every baseline, using the double-difference
ionospheric free residuals for 16 and 17 March, every 30 s. Figure A.4A, modified
from Stolk (1997), shows an example of the relative troposphere estimations for
day 76, baseline Assen-Blijham. A considerable amount of noise is apparent in these
data, making a direct comparison of the raw 30 s GPS data with the interferogram
impossible. The result of low-pass filtering the data, with a cut-off frequency of
is shown overlaid on fig. A.4A. Unfortunately, the temporal variation
in the tropospheric delay of day 76 for the 33 km baseline Assen-Blijham is very
limited, less than 2 cm.
Using the low-pass filter for all baselines for both days enables the extraction of the
relative tropospheric wet delay values during the two SAR acquisitions. These values
for both days are then subtracted to mimic the difference between the two days in
the interferogram. Using a reference station, where the values for both methods
are assigned equal (null) values, the absolute difference in tropospheric zenith delay
observations between both acquisition times is determined at every GPS station.
This is shown in fig. A.4B. On the horizontal axes the location of the receivers
in azimuth direction is plotted. The error bars represent a 95% confidence
interval. Although it is obvious that the atmospheric signal during this experiment is
limited, a first quantitative evaluation between the InSAR and the GPS results shows
agreement between both methods: the observed rms of the difference is ~3 mm.
Recognizing the possibility of a phase trend in the interferogram, and solving for
262 Appendix A: Comparison neutral delay GPS and InSAR

such a first-order trend reduces the rms of the difference to ~2 mm. In fig. A.6, the
comparison between the six GPS baselines with respect to station Slochteren and
the SAR data is shown. The location of the points, along the diagonal, suggests
a correlation between both data sources, although the variances of the data are
relatively large.
The possibility to “correct” the interferometric data using the (interpolated) GPS
data is investigated and shown in fig. A.5. Figure A.5A shows the interferogram,
which might include a linear phase trend due to orbit errors. In the corresponding
histogram of the phase values, the interferogram including trend is shown in gray,
while the interferogram corrected for a bilinear phase trend is shown in black. The
GPS data, see fig. A.5B, are interpolated using an isotropic harmonic spline, with
local maxima/minima occurring only at the data points (Wessel and Smith, 1998).
It is obvious that the extrapolation of the data outside the GPS station framework
can be erroneous, although it could as well be an indication for a true trend in the
atmospheric delay. There is no way to solve this problem without any additional
data, although a significant trend in the delay is not likely, considering the weather
conditions. The histogram corresponding to the GPS data reflects the trend in the
data. Eliminating the trend in the InSAR data by fitting it to the GPS data yields
a “corrected” result. In this result, the interferogram adopts the GPS trend and
should be corrected for atmospheric signal with, say, 10–20 km wavelengths, based
on the distance between the receivers. In fig. A.5C, we subtract the trend, which
would disturb the visual comparison and cannot be validated, and show only the
residual signal for comparison. Evaluation of the histograms for the detrended data,
in fig.A.5A and C, shows no significant improvement of the data distribution.

A.1.4 Discussion

Due to the fact that the amount of spatial variability in the atmospheric delay was
extremely limited during this specific experiment, the methodological validation of
an atmospheric correction of interferograms using GPS networks is difficult. More-
over, the L3 relative delays in the GPS baselines are too noisy for a direct comparison
with the interferogram, resulting in error margins that exceed the atmospheric delay
variability in this situation, even though there is correlation between both datasets,
see fig. A.6. Therefore, it is recommended to use L1 and/or L2 processing instead
of the ionospheric-free (L3) linear combination.
Low-pass filtering the data (effectively averaging over periods of ~80 minutes) re-
duces the noise, but is rather arbitrary and crude. Therefore, the results obtained
from this experiment do not allow for general statements on the feasibility of cor-
rection of radar interferograms using a GPS network. Further experiments, during
variable weather conditions, are necessary to achieve better understanding of the
possibilities. Nevertheless, based on this experiment we can comment on some im-
portant issues.
The possibility for correction of the interferograms is dependent on the spatial distri-
bution of the GPS receivers. Since (continuous) GPS networks are usually designed
for other purposes, sampling criteria for SAR purposes are not optimal, and the re-
 A.1 Spatial networks 263

duction of the atmospheric noise using interpolation between the receivers is mainly
achieved over areas directly surrounding the GPS receivers. On the other hand,
considering a densification of a GPS network there are limitations related to the
zenith delay reduction of the data. Using a 20-degree elevation angle, the zenith
delay estimate will reflect atmospheric variation within a wide range, and cannot be
considered as a point measurement, see fig.A.2. For atmospheric signal at 1, 2, or 3
km altitude, this elevation angle results in circles with a diameter of 5, 11, and 16 km
respectively. Therefore, the zenith delay will reflect some kind of weighted average
of all atmospheric variation within that area, and it will not be possible to remove
atmospheric artifacts with smaller scales from the interferogram. This results in an
upper limit to the effective density of the GPS stations, in effect a smoothing of the
signal whenever the GPS receivers are too close. This is discussed in more detail in
appendix A.2.2. Therefore, future studies need to address the problem of slant-delay
measurements instead of zenith-delay measurements.
For deformation studies, there is another, more fundamental, reason that limits the
264 Appendix A: Comparison neutral delay GPS and InSAR

value of GPS networks for atmospheric corrections in SAR interferograms. With a

continuous GPS network it is possible to measure the long wavelength displacement
field in a region which may be deforming. Unfortunately, deformations on small
spatial scales cannot be observed, due to the limitations in the receiver density. Fil-
tering the long wavelength atmospheric signal from a SAR interferogram still leaves
all atmospheric (and deformation) signal on these small spatial scales in the image.
In fact, the technique even introduces ambiguous short wavelength signal due to the
limitations in the interpolation. As a result, for short wavelengths the interpretation
of a GPS-filtered interferogram can be even more difficult than without the filtering,
since it is a superposition of atmospheric and interpolation errors. On the other
hand, for long wavelengths, the deformation signal was already known from GPS,
which makes SAR superfluous.

A.2 Single-station time-series

A single GPS station has been selected to investigate the situation when no dense
network of GPS receivers is available for the observation period. Unfortunately, this
is common practice for most of the situations where research with SAR images is
involved. Moreover, although a dense (permanent) network of GPS receivers would
enable a direct comparison of the spatial delay variability, there are considerable
differences as well. Limitations in this approach are (i) the weighted averaging to
zenith delays and (ii) the introduction of interpolation errors due to the inhomoge-
neous distribution of GPS receivers. If such techniques are applied “blindly” and
used for extracting an atmospheric phase screen from the SAR interferograms, the
geophysical interpretation of the results may be more ambiguous than without such
a correction, as demonstrated in the previous section.
 A.2 Single-station time-series 265

This experiment is not an attempt to correct the interferograms, but rather to ob-
tain reasonable estimates of the delay variations and spatial scales to be expected.
Such quantitative estimates could significantly improve the interpretation of inter-
ferometric SAR data, using only a single GPS receiver as additional source of infor-
mation (van der Hoeven and Hanssen, 1999).
Five case-studies are performed, using ERS-1 and ERS-2 SAR data acquired with
a one-day interval, hereby excluding crustal deformation signal. Swaths of 200 ×
100 km are processed differentially, removing the topographic signal using a reference
elevation model (TDN/MD, 1997), see section 2.4.2. After applying a simple cosine
mapping function, the residual signal consists of zenith delay variations in each
pair of data acquisitions (Hanssen et al., 1999). GPS data are obtained at the
Kootwijk Observatory for Satellite Geodesy (KOSG), a permanent GPS receiving
station. The GPS observations are processed using GIPSY-OASIS while solving for
a free network. Data from 13 widely spread IGS stations are used to solve for a
number of parameters including satellite clocks, station positions, gradients, and
tropospheric delays. The receiver clocks were estimated as white noise processes,
while the wet tropospheric delays were estimated as random-walk processes, using
the Niell mapping function to convert slant delays from a minimum elevation of 10o
to zenith delays (Niell, 1996). The a priori hydrostatic delay is calculated using the
Saastamoinen model. To reduce the noise in the zenith delay estimates, the data
obtained at a sampling rate of 30 s are averaged into 6 min intervals. The fluctuations
in the zenith delays consist of the combined hydrostatic and wet delay—the same
contributions as measured by InSAR.

A.2.1 Methodology of comparison

The absolute zenith tropospheric delays derived from GPS constitute time-series
at a fixed position. In contrary, InSAR generates a spatial image of the relative
zenith delays, differenced between two fixed acquisition times. A conversion needs
to be performed to connect the relative-spatial and absolute-temporal observations
in order to validate the two datasets.
Two assumptions have been applied to match the two techniques. First, Taylors
approximation is applied, treating the local atmosphere as a frozen atmosphere
moving over the area without changing its refractivity distribution during a preset
time interval (Taylor, 1938; Treuhaft and Lanyi, 1987). This refractivity distribu-
tion is displaced by the mean wind speed and direction, obtained from surface or
radiosonde observations. Second, wind speed and wind direction are assumed to
be approximately equal during both observation days. The latter assumption will
certainly fail for longer time intervals, but can be more likely for a one-day interval,
a common choice for, e.g., DEM generation with InSAR.
Applying these assumptions the GPS time-series can be converted to spatial zenith
delay profiles, corresponding with a ground trace in the wind direction and stretched
by the wind velocity. These profiles are computed for the two SAR observations from
about 4 hours before until about 2 hours after the SAR acquisition (21.41 UTC).
The non-symmetric epoch was chosen to avoid delay errors caused by discontinuities
266 Appendix A: Comparison neutral delay GPS and InSAR

in the GPS satellite orbit estimation at the day break, a common problem in con-
temporary GPS processing. Subsequently the two profiles are differenced, resulting
in a differential delay profile, which is comparable to a cross-section in the interfero-
gram. The arbitrary bias between the two sources of data is removed by subtracting
their means. This approach is referred to as the synoptic approach.
To verify the feasibility of using synoptic wind observations, an alternative approach
is performed by rotating and stretching the ground trace over the InSAR-image while
recording the correlation between the two profiles. The GPS observations made at
the time of the SAR observations are used as axis of rotation. Maximum correlation
 A.2 Single-station time-series 267

corresponds with the best-fit between GPS and InSAR. The length of the best-fit
ground trace divided by the total GPS observation time is a measure for the wind
speed. The angle of rotation of the best-fit ground trace gives the effective wind
direction relative to the local north. The derived wind speed and direction can then
be compared with the synoptic observations for validation.

A.2.2 Results

The results of the tropospheric delays estimated by GPS and InSAR are shown in
fig. A.7 and table A.1. The first row of fig. A.7 shows the five differential inter-
ferograms together with the estimated best-fit profile in black-blue, and the profile
based on the synoptic observations in black-red. The color changes in the profiles
represent 1 hour time intervals. The location of the GPS receiver is indicated by the
red star. All interferograms have been converted to relative zenith delay differences,
expressed in mm. One colorbar is used, to enable the comparison of the magnitude
in variation in every case.
The profiles below the interferograms correspond with the GPS observations (black
dots), the scaled SAR profile which shows a best fit with the GPS time series (blue),
and the SAR profile which is obtained using the surface wind speed and direction
(red). In the lower row of plots, correlation coefficients are depicted which resulted
in the best-fit profile. The vertical axis corresponds with the direction of the profile,
the horizontal axis scales the profile with the wind speed in m/s.
Synoptic data from a meteo-station 30 km north of KOSG were used. Average wind
speeds were calculated using the data from 21.00 and 22.00 UTC for both days.
268 Appendix A: Comparison neutral delay GPS and InSAR

When the average wind speed was below 2.5 m/s it was assumed that no significant
delay changes were present and so this synoptic data was not used in the averaging.
Table A.1 lists, for every interferogram. the correlation coefficient between the pro-
files and the interferometric data. The wind speed and wind direction can be derived
based on the amount of rotation and stretching of the GPS profile. This can be com-
pared with the observed synoptic wind speeds. The comparison between GPS and
SAR yields an rms of difference, which can be compared to the total range of delay
variation.

A.2.3 Discussion

The five analyzed interferograms represent very different weather situations, ranging
from large-scale humidity variation, a narrow and a wide cold front, to a relatively
 A.2 Single-station time-series 269

undisturbed, well-mixed boundary layer. Nevertheless, relatively high correlations

are found between the GPS and the SAR data. Table A.1 shows that the best-fit
correlations between GPS and InSAR for the five analyzed interferograms are 0.8 or
better, while the estimated wind directions are accurate to within 30deg and wind
speeds differ maximally 4.0 m/s. In case the synoptic data are used to situate the
SAR profile, correlations are 0.75 or better, except for the May 1996 interferogram,
where a correlation of –0.01 is found.
RMS values between GPS and the synoptic SAR profiles are generally worse than
those between GPS and the best-fit SAR profiles, which is also indicated by the
correlation values. This can be due to (i) a difference in wind speed and wind
direction between the two days, and (ii) the difference in wind speed and direction
at surface level and aloft. Furthermore, the assumption of a frozen atmosphere might
not be valid in all cases or limited in time.
The sensitivity of GPS observations and SAR interferograms for horizontal refractiv-
ity variations at different scales is dependent on the spatial sampling characteristics,
the data processing strategy, and the measurement accuracy. Using the frozen at-
mosphere model with constant wind speed and direction, GPS time series can be
regarded as spatial observations along a straight line, with a sampling rate deter-
mined by the wind speed. Figure A.8 is a conceptual sketch of the sensitivity of
GPS and InSAR for atmospheric variation at different wavelengths.
For InSAR, the sensitivity ranges vary theoretically from the resolution cell size to
the size of the interferogram. Since measurement noise affects the phase measure-
ments at single resolution cells considerably, an averaging to 100–200 m is used to
suppress this noise. For large wavelengths, orbit inaccuracies may cause nearly lin-
ear trends in the interferogram, see section 4.6. Both effects limit the effective part
of the spectrum to wavelengths between 50 km and 100 m.
A GPS network approach is independent of wind speed, assuming the zenith delays
can be obtained instantaneous. Effectively, the lower part of the spectrum is not
limited, as it is dependent of the amount of receivers and the spatial extent of
the network. If the system would be able to measure true zenith delays, the short
wavelength part of the spectrum would simply be determined by the spacing between
the receivers. Unfortunately, accurate zenith delay estimates are currently only
obtained using a spatial and temporal averaging procedure over many satellites
above an a priori defined elevation cut-off angle, see the inset at fig. A.8. This
procedure effectively acts as a low-pass filter on the spatial wavelengths, with a cut-
off wavenumber determined by the minimum elevation angle and the scale height
of the wet troposphere, below which most of the horizontal refractivity variations
occur. This scale is indicated in the figure as the cone diameter limit.
The use of GPS time series, as described above, implies a dependency of the wind
speed. The horizontal lines in fig. A.8 indicate situations in which the wind speed
is either too low or too high for the assumption of a frozen atmosphere to be valid.
The left-most vertical line shows that for large spatial scales it will take very long for
the atmosphere to drift over the GPS receiver—too long to regard it as frozen. The
conversion from temporal GPS observations with an effective sampling frequency of
270 Appendix A: Comparison neutral delay GPS and InSAR

to a spatial sampling, as a function of wind speed yields a spatial

Nyquist wavenumber of

indicated by the curved line in fig. A.8. It is shown that for wind speeds higher
than ~7 m/s, the data are undersampled, which might result in aliasing effects.
On the other hand, for lower wind speeds the data are effectively oversampled and
can be regarded as bandlimited. Low-pass filtering should be applied to suppress
short wavelengths higher than the bandwidth determined by the cone diameter. The
effective range of wavelengths and wind speeds for which GPS time series can be
applied is indicated by the shaded region in fig. A.8.

A.2.4 Conclusions

When comparing water vapor time-series from a single GPS station with water vapor
profiles derived from SAR interferograms it is necessary to have comparable wind
speeds and wind directions on both observation days. For the mentioned cases,
where the wind speed doesn’t differ more than 30 degrees between both days, and
the wind speed is equal within 10%, correlations are found in the range of 0.91-0.98.
During the situations with larger differences, 30 degrees and 50% respectively, the
correlations decreased to 0.81-0.85. This technique only works for a limited strip
of the interferogram, i.e., the area where the measured time series covers the SAR
image. Nevertheless, the study shows a clear and strong correspondence between
wet signal delay as observed by GPS and InSAR. As such, it aids the interpretation
of GPS studies, recognizing the spatial variability of the wet delay and the validity of
assumptions on homogeneous or gradient atmospheric models. Currently, the GPS
data are used in experiments to parameterize the stochastic model of atmospheric
signal for SAR interferometry.
 Appendix B

Structure function and power spectrum

This appendix elaborates on the relation between the structure function of the
refractivity and the structure function of the (integrated) signal delay, as used
in chapter 4. In the second part, the analogy between the structure function
and the power spectrum is discussed.
key words: Structure function, Power spectrum, Refractivity, Signal
delay

B.1 Structure function for refractivity and signal delay

It can be shown that the structure function of the delay see eq. (4.7.6), can
be related to the structure function of the refractivity eq. (4.7.1) (Tatarski,
1961; Treuhaft and Lanyi, 1987; Coulman and Vernin, 1991):

with and

and

The geometric configuration used in this equation is sketched in fig. B.1.

272 Appendix B: Structure function and power spectrum

Derivation. Starting from the definition of the structure function in eq. (4.7.6) we
substitute eq. (4.7.5):

First we decompose the quadratic term. As the expectation of a sum is equal to the
sum of the expectations, we can exchange integration and expectation (or ensemble
average):

We removed the (horizontal) term, assuming that E{N} is independent of the

horizontal position. Writing the integral outside of the summation, we can simplify
eq. (B.1.5) to
 B.2 Structure function and power-law spectrum 273

For second-order stationary signals we can write the structure function of N, eq. (4.7.1),
as

which yields an expression of the structure function of the delay as a function of the
structure function of the refractivity N:

as in eqs. (B.1.1) and (4.7.7).

End of derivation

B.2 Structure function and power-law spectrum

The structure function of a function is defined as

where a new signal is defined as

To relate the structure function to the power-law spectral form of eq. (4.7.12),
we can rewrite in eq. (B.2.2) in a more general form using the convolution
notation, see fig. B.2 (Bracewell, 1986):

where the act as a spatial filter. The power spectrum of can be

written as the product of two power spectra:
274 Appendix B: Structure function and power spectrum

where is the power spectrum of and the power spectrum of

with

Using and using the shift theorem (Bracewell, 1986), we find

and using the linearity conditions of Fourier transforms we obtain

Using we can retrieve the power spectrum by

complex multiplication

Hence, the power spectrum of see eq. (B.2.4) is given by

The variance of a random process is equal to the integral of its power spectrum
over all frequencies, using eqs. (B.2.8) and (B.2.4):

If we now insert the special power spectral form of (4.7.12), we can write (B.2.10)
as

Substituting this can be written as

 B.2 Structure function and power-law spectrum 275

or. as in Agnew (1992)

We use the following definite integral (Gradshteyn et al., 1994)

with and This results in

From the recurrence relation, see, e.g., (Arfken, 1985) or (Monin et al., 1975)

we derive

or

and using we find

Including (B.2.18) in (B.2.15) yields

Using and by rearranging terms we finally find

or

as in eq. (4.7.13).
End of derivation.
 Bibliography

Abramowitz, M. and Stegun, I. A. (1970), Handbook of Mathematical Functions, Dover

books on advanced mathematics, Dover, New York, 9th edn.
Adam, N., Eineder, M., Breit, H. and Suchandt, S. (1999), Shuttle Radar Topography
Mission: DLR’s Interferometric SAR Processor for the Generation of a Global Digi-
tal Elevation Model, in: Second International Workshop on ERS SAR Interferometry,
‘FRINGE99’, Liège, Belgium, 10–12 Nov 1999, ESA.
Afraimovich, E. L., Terekhov, A. I., Udodov, M. Y. and Fridman, S. V. (1992), Refraction
distortions of transionospheric radio signals caused by changes in a regular ionosphere and
by travelling ionospheric disturbances, Journal of Atmospheric and Terrestrial Physics,
54(7/8):1013–1020.
Agnew, D. C. (1992), The time-domain behaviour of power-law noises, Geophysical Research
Letters, 19(4):333–336.
Ahmed, S., Warren, H. R., Symonds, D. and Cox, R. P. (1990), The Radarsat system,
IEEE Transactions on Geoscience and Remote Sensing, 28:598–602.
Alaska SAR Facility (1997), ERS-1 and ERS-2 SAR Images, ASF, (Accessed 22 March,
1998), <http://www.asf.alaska.edu/>.
Alpers, W. and Brümmer, B. (1994), Atmospheric boundary layer rolls observed by the
synthetic aperture radar aboard the ERS-1 satellite, Journal of Geophysical Research,
99(C6):12613–12621.
Amelung, F., Galloway, D. L., Bell, J. W., Zebker, H. A. and Laczniak, R. J. (1999), Sensing
the ups and downs of Las Vegas: InSAR reveals structural control of land subsidence
and aquifer-system deformation, Geology, 27(6):483–486.
Amelung, F., Jónssen, S., Zebker, H. and Segall, P. (2000a), Widespread uplift and
trap door faulting on Galapagos volcanoes observed with radar interferometry, Nature,
407(6807) :993–996.
Amelung, F., Oppenheimer, C., Segall, P. and Zebker, H. (2000b), Ground deformation
near Gada ’Ale Volcano, Afar, observed by radar interferometry, Geophysical Research
Letters, 27(19):3093–3097.
Ammar, G. S. and Gragg, W. B. (1988), Superfast solution of real positive definite Toeplitz
Systems, SIAM J. Matrix Anal. Appl,, 9(l):61–76.
Arfken, G. (1985), Mathematical Methods for Physicists, Academic Press, Boston, 3rd edn.
Armstrong, J. W. and Sramek, R. A. (1982), Observations of tropospheric phase scintilla-
tions at 5 GHz on vertical paths, Radio Science, 17(6): 1579–1586.
Askne, J. and Hagberg, J. O. (1993), Potential of Interferometric SAR for Classification
of Land Surfaces, in: International Geoscience and Remote Sensing Symposium, Tokyo,
Japan, 18–21 August 1993, pp. 985–987.
Askne, J. and Nordius, H. (1987), Estimation of tropospheric delay for microwaves from
278 Bibliography

surface weather data, Radio Science, 22:379–386.

Atkinson, B, W., and Zhang, J. W. (1996), Mesoscale shallow convection in the atmosphere,
Reviews of Geophysics, 34(4):403–431.
Attema, E. P. W. (1991), The Active Microwave Instrument On-Board the ERS-1 Satellite,
Proceedings of the IEEE, 79(6):791–799.
Avallone, A., Zollo, A., Briole, P., Delacourt, C. and Beauducel, F. (1999), Subsidence
of Campi Flegrei (Italy) detected by SAR interferometry, Geophysical Research Letters,
26(15):2303–2306.
Baby, H. B., Golé, P. and Lavergnat, J. (1988), A model for the tropospheric excess
path length of radio waves from surface meteorological measurements, Radio Science,
23(6):1023–1038.
Baer, G., Sandwell, D., Williams, S., Bock, Y. and Shamir, G. (1999), Coseismic defor-
mation associated with the November 1995, Nuweiba earthquake, Gulf of Elat
(Aqaba), detected by synthetic aperture radar interferometry, Journal of Geophysical
Research, 104(B11):25221–25232.
Balakrishnan, A. V. (1995), Introduction to random processes in engineering, John Wiley,
New York.
Baltsavias, E. (1993), Fotogrammetrie III, Lecture notes, Delft University of Technology,
Faculty of Geodetic Engineering.
Bamler, R. (1992), A comparison of range-Doppler and wave-number domain SAR focusing
algorithms, IEEE Transactions on Geoscience and Remote Sensing, 30(4):706–713.
Bamler, R., Adam, N., Davidson, G. W. and Just, D. (1998), Noise-Induced Slope Distor-
tion in 2-D Phase Unwrapping by Linear Estimators with Application to SAR Interfer-
ometry, IEEE Transactions on Geoscience and Remote Sensing, 36(3):913–921.
Bamler, R., Davidson, G. W. and Adam, N. (1996a), On the Nature of Noise in 2-D Phase
Unwrapping, in: Franceschetti et al. (1996).
Bamler, R., Eineder, M. and Breit, H. (1996b), The X-SAR Single-Pass Interferometer on
SRTM: Expected Perfomance and Processing Concept, in: EUSAR’96, Königswinter,
Germany, pp. 181–184.
Bamler, R. and Hartl, P. (1998), Synthetic aperture radar interferometry, Inverse Problems,
14:R1–R54.
Bamler, R. and Just, D. (1993), Phase Statistics and Decorrelation in SAR Interferograms,
in: International Geoscience and Remote Sensing Symposium, Tokyo, Japan, 18–21 Au-
gust 1993, pp. 980–984.
Bamler, R. and Schättler, B. (1993), SAR Data Acquisition and Image Formation, in:
Schreier, G., ed., SAR Geocoding: data and systems, pp. 53–102, Wichmann Verlag,
Karlsruhe.
Barber, B. C. (1993a), The non-isotropic two-dimensional random walk, Waves in Random
Media, 3:243–256.
Barber, B. C. (1993b), The phase statistics of a multichannel radar interferometer, Waves
in Random Media, 3:257–266.
Barbieri, M., Lichtenegger, J. and Calabresi, G. (1999), The Izmit Earthquake: A Quick
Post-Seismic Analysis with Satellite Observations, ESA Bulletin, 100:107–110.
Barka, A. (1999), The 17 August 1999 Izmit Earthquake, Science, 285:1858–1859.
Bean, B. R. and Dutton, E. J. (1968), Radio Meteorology, Dover, New York.
Bendat, J. S. and Piersol, A. G. (1986), Random Data: Analysis and Measurement Proce-
dures, Wiley-Interscience, New York, 2nd edn.
Bennett, J. R. and Cumming, I. G. (1979), A Digital Processor for the Production of Seasat
Synthetic Aperture Radar Imagery, Tech. Rep. ESA SP-154, ESA.
Bevington, P. R. and Robinson, D. K. (1992), Data reduction and analysis for the physical
 Bibliography 279

sciences, McGraw-Hill, Boston, 2nd edn.

Bevis, M., Businger, S., Herring, T. A., Anthes, R. A., Rocken, C. and Ware, R. H. (1994),
GPS meteorology: mapping zenith wet delays onto precipitable water, Journal of Applied
Meteorology, 33(3).
Bevis, M., Businger, S., Herring, T. A., Rocken, C., Anthes, R. A. and Ware, R. H.
(1992), GPS Meteorology: Remote sensing of atmospheric water vapor using the Global
Positioning System, Journal of Geophysical Research, 97:15,787–15,801.
Bevis, M., Chiswell, S., Businger, S., Herring, T. A. and Bock, Y. (1996), Estimating wet
delays using numerical weather analysis and predictions, Radio Science, 31(3):477–487.
Bitmead, R. R. and Anderson, B. D. O. (1980), Asymptotically Fast Solution of Toeplitz
and Related Systems of Linear Equations, Linear Algebra and Its Applications, 34:103–
116.
Bochner, S. (1959), Lectures on Fourier Integrals; with a supplement on monotonic func-
tions, Stieltjes integrals, and harmonic analysis;, Annals of mathematics studies 42,
Princeton University Press, Princeton, transl. from German by M. Tenenbaum and
H. Pollard.
Born, M. and Wolf, E. (1980), Electromagnetic Theory of Propagation Interference and
Diffraction of Light, Pergamon Press, New York.
Born, M., Wolf, E. and Bhatia, A. B. (1959), Principle of optics; electromagnetic theory of
propagation, interference and diffraction of light, Pergamon Press, New York.
Box, G. E. P., Jenkins, G. M. and Reinsel, G. C. (1994), Time series analysis: forecasting
and control, Prentice-Hall, Englewood Cliffs, 3rd edn.
Bracewell, R. N. (1986), The Fourier transform and its applications, McGraw-Hill, New
York, 3rd edn.
van Bree, R. J. P., Gens, R., Groot, J. S., van Halsema, D., Hanssen, R. F., van den
Hout, P., Klees, R., de Min ans E J O Schrama, E. J. and Usai, S. (2000), Deformation
measurements with SAR, vol. USP-2, 99-16, Beleidscommissie Remote Sensing, Delft.
Brent, R. P. (1991), Parallel Algorithms for Toeplitz Systems, in: Golub, G. G. and Dooren,
P. V., eds., Numerical Linear Algebra, Digital Signal Processing and Parallel Algorithms,
vol. 70 of NATO ASI Series, Series F: Computer and Systems Sciences, pp. 75–92,
Springer-Verlag, Berlin.
Briole, P., Massonnet, D. and Delacourt, C. (1997), Post-eruptive deformation associated
with the 1986-87 and 1989 lava flows of Etna detected by radar interferometry, Geophys-
ical Research Letters, 24:37–40.
Brown, L. G. (1992), A Survey of Image Registration Techniques, ACM Computing Surveys,
24(4):325–376.
Brown, R. A. (1980), Longitudinal instabilities and secondary flows in the planetary bound-
ary layer: A review, Reviews in Geophysics and Space Physics. 18:683–697.
Brunner, F. K. and Welsch, W. M. (1993), Effect of the Troposphere on GPS Measurements,
GPS World, pp. 42–51.
Buderi, R. (1996), The invention that changed the world, Simon & Schuste, New York.
Bürgmann, R., Schmidt, D., Nadeau, R. M., d’Alessio, M., Fielding, E. J., Manaker, D.,
McEvilly, T. V. and Murray, M. H. (2000), Earthquake Potential Along the Northern
Hayward Fault, California, Science, 289:1178–1182.
Calais, E. and Minster, J. B. (1995), Detection of Ionospheric Perturbations Following
an Earthquake with the Global Positioning System: Implications for Nuclear Tests Dis-
crimination, in: Console, R. and Nikolaev, A., eds., Earthquakes Induced by Underground
Nuclear Explosions, vol. 4 of NATO ASI Series, Partnership Subseries, 2. Environment,
pp. 235–253, Springer-Verlag, Berlin.
Calais, E. and Minster, J. B. (1996), GPS detection of ionospheric perturbations following
280 Bibliography

a Space Shuttle ascent, Geophysical Research Letters, 23(15):1897–1900.

Calais, E. and Minster, J. B. (1998), GPS, earthquakes, the ionosphere, and the Space
Shuttle, Physics of the Earth and Planetary Interiors, 105:167–181.
Carnec, C. and Fabriol, H. (1999), Monitoring and modeling land subsidence at the Cerro
Prieto geothermal field, Baja California, Mexico, using SAR interferometry, Geophysical
Research Letters, 26(9):1211–1215.
Carnec, C., Massonnet, D. and King, C. (1996), Two examples of the use of SAR inter-
ferometry on displacement fields of small spatial extent, Geophysical Research Letters,
23(24):3579–3582.
Carrasco, D., Alonso, J. and Broquetas, A. (1995), Accuracy assessment of SAR interfer-
ometry using the ERS-1, in: International Geoscience and Remote Sensing Symposium,
Florence, Italy, 10–14 July 1995, pp. 781–783.
Cattabeni, M., Monti-Guarnieri, A. and Rocca, F. (1994). Estimation and Improvement
of Coherence in SAR Interferograms, in: International Geoscience and Remote Sensing
Symposium, Pasadena, CA, USA, 8–12 August 1994, PP. 720–722.
Chen, C. W. and Zebker, H. A. (2000a), Network approaches to two-dimensional phase
unwrapping: intractability and two new algorithms, Journal of the Optical Society of
America A., 17(3):401–414.
Chen, C. W. and Zebker, H. A. (2000b), Two-dimensional phase unwrapping using statis-
tical models for cost functions in nonlinear optimization, Journal of the Optical Society
of America A., in press.
Cheng, K. and Huang, Y.-N. (1992), Ionospheric disturbances observed during the pe-
riod of Mount Pinatubo eruptions in June 1991, Journal of Geophysical Research,
97(A11):16,995–17,004.
Chilès, J.-P. and Delfiner, P. (1999), Geostatistics: modeling spatial uncertainty, Wiley
Series in Probability and Statistics, John Wiley & Sons, New York.
Ching, N. H., Rosenfeld, D. and Braun, M. (1992), Two-dimensional phase unwrapping
using a minimum spanning tree algorithm, IEEE Transactions on Image Processing,
1:355–365.
Clarke, P. J., Paradissis, D., Briole, P., England, P. C., Parsons, B. E., Billiris, H., Veis,
G., and Ruegg, J.-C. (1996), Geodetic investigation of the 13 May 1995 Kozani-Grevena
(Greece) earthquake, Geophysical Research Letters, 24:707–710.
Coltelli, M., Fornaro, G., Franceschetti, G., Lanari, R., Migliaccio, M., ao R Moreira,
J., Papatanassiou, K. P., Puglisi, G., Riccio, D. and Schwäbisch, M. (1996), SIR-C/X-
SAR multifrequency multipass interferometry: A new tool for geological interpretation,
Journal of Geophysical Research, 101(E10):23127–23148.
Costantini, M. (1996), A Phase Unwrapping Method Based on Network Programming,
in: ‘FRINGE 96’ workshop on ERS SAR Interferometry, Zurich, Switzerland, 30 Sep–2
October 1996.
Costantini, M. (1998), A Novel Phase Unwrapping Method Based on Network Program-
ming, IEEE Transactions on Geoscience and Remote Sensing, 36(3):813–821.
Coulman, C. E. and Vernin, J. (1991), Significance of anisotropy and the outer scale of
turbulence for optical and radio seeing, Applied Optics, 30(1):118–126.
Coulson, S. N. (1995), SAR Interferometry with ERS-1.
Crook, N. A. (1996), Sensitivity of moist convection forced by boundary layer processes to
low-level thermodynamic fields, Monthly Weather Review, 124:1767–1785.
Curlander, J. C. and McDonough, R. N. (1991), Synthetic aperture radar: systems and
signal processing, John Wiley & Sons, Inc, New York.
Dainty, J. C., ed. (1975), Laser Speckle and Related Phenomena, vol. 9 of Topics in Applied
Physics, Springer-Verlag, Heidelberg.
 Bibliography 281

Dammert, P. B. G. (1997), Accuracy of INSAR Measurements in Forested Areas, in:

‘FRINGE 96’ workshop on ERS SAR Interferometry, Zurich, Switzerland, 30 Sep–2
October 1996, ESA SP-406, Vol II.
Daniels, M. J. and Cressie, N. (1999), A hierarchical approach to covariance function
estimation for time series, submitted to Journal of Time Series Analysis.
Davenport, Jr, W. B. and Root, W. L. (1987), An Introduction to the Theory of Random
Signals and Noise, IEEE Press, New York.
Davis, A., Marshak, A., Wiscombe, W. and Cahalan, R. (1994), Multifractal characteriza-
tion of nonstationarity and intermittency in geophysical fields: Observed, retrieved, or
simulated, Journal of Geophysical Research, 99(D4):8055–8072.
Davis, J. L., Herring, T. A., Shapiro, I. I., Rogers, A. E. E. and Elgered, G. (1985), Geodesy
by radio interferometry: Effects of atmospheric modelling errors on estimates of baseline
length, Radio Science, 20(6): 1593–1607.
Delacourt, C., Briole, P. and Achache, J. (1998), Tropospheric corrections of SAR inter-
ferograms with strong topography. Application to Etna, Geophysical Research Letters,
25(15):2849–2852.
Delacourt, C., Briole, P., Achache, J., Fruneau, B. and Carnec, C. (1997), Correction of the
tropospheric delay in SAR interferometry and application to 1991-93 eruption of Etna
volcano, Italy, in: AGU Fall meeting, December 8-12, San Francisco, USA,.
Delaunay, B. N. (1934), Sur la sphère vide, Izvestia Akademia Nauk SSSR, VII Seria,
Otdelenie Matematicheskii i Estestvennyka Nauk, 7:793–800.
Dravskikh, A. F. and Finkelstein, A. M. (1979), Tropospheric limitations in phase and fre-
quency coordinate measurements in astronomy, Astrophysics and Space Science, 60:251–
265.
Duchossois, G., Kohlhammer, G. and Martin, P. (1996), Completion of the ERS Tandem
Mission, Earth Observation Quarterly, 55.
Elachi, C. (1988), Spaceborne radar remote sensing: applications and techniques, Institute
of Electrical and Electronics Engineers, New York.
Elachi, C., Bicknell, T., Jordan, R. L. and Wu, C. (1982), Spaceborne Synthetic-Aperture
Imaging Radar: Applications, Techniques and Technology, Proceedings of the IEEE,
70(10):1174–1209.
Elgered, G. (1982), Tropospheric wet path-delay measurements, IEEE Trans. on Antennas
and Propagation, AP-30:502–505.
Emanuel, K., Raymond, D., Betts, A., Bosart, L., Bretherton, C., Droegemeier, K., Farrell,
B., Fritsch, J. M., Houze, R., LeMone, M., Lilly, D., Rotunno, R., Shapiro, M., Smith,
R. and Thorpe, A. (1995), Report of the first prospectus development team of the U. S.
weather research program to NOAA and the NSF, Bull. Amer. Meteor. Soc., 76:1194–
1208.
Emardson, T. R. and Derks, H. J. P. (2000), On the relation between the wet delay and
the integrated precipitable water vapour in the European atmosphere, Meteorological
Applications, 7(l):61–68.
European Space Agency (1999), Active Microwave Instrument, (Accessed 13 Aug, 1999)
<http://earthl.esrin.esa.it/ipgami>.
Evans, D. L., Elachi, C., Stofan, E. R., Holt, B., Way, J. B., Kobrick, M., Öttl, H.,
Pampaloni, P., Vogt, M., Wall, S., van Zyl, J. and Schier, M. (1993), The Shuttle
Imaging Radar-C and X-SAR Mission, EOS, 74(13).
Evans, J. V. and Hagfors, T. (1968), Radar Astronomy, McGraw-Hill, New York.
Fabriol, H. and Glowacka, E. (1997), Seismicity and fluid reinjection at CerroPrieto geother-
mal field: preliminary results, in: Twenty-Second Workshop on Geothermal Reservoir
Engineering, pp. 11–17, Stanford University, Jan 27-29, 1997.
282 Bibliography

Feigl, K. L., Sergent, A. and Jacq, D. (1995), Estimation of an earthquake focal mecha-
nism from a satellite radar interferogram: Application to the December 4, 1992 Landers
aftershock, Geophysical Research Letters, 22(9):1037–1040.
Feijt, A. J. (2000), Quantitative Cloud Analysis using Meteorological Satellites, Ph.D. thesis,
Wageningen University.
Ferretti, A., Monti-Guarnieri, A., Prati, C. and Rocca, F. (1998), Multi-image DEM Recon-
struction, in: International Geoscience and Remote Sensing Symposium, Seattle, Wash-
ington, USA, 6–10 July 1998, pp. 1367–1369.
Ferretti, A., Prati, C. and Rocca, F. (1999a), Multibaseline InSAR DEM Reconstruc-
tion: The Wavelet Approach, IEEE Transactions on Geoscience and Remote Sensing,
37(2):705–715.
Ferretti, A., Prati, C. and Rocca, F. (1999b), Non-Uniform Motion Monitoring Using
the Permanent Scatterers Technique, in: Second International Workshop on ERS SAR
Interferometry, ‘FRINGE99’, Liège, Belgium, 10–12 Nov 1999, pp. 1–6, ESA.
Ferretti, A., Prati, C. and Rocca, F. (2000), Nonlinear Subsidence Rate Estimation using
Permanent Scatterers in Differential SAR Interferometry, IEEE Transactions on Geo-
science and Remote Sensing, 38(5):2202–2212.
Ferretti, A., Prati, C., Rocca, F. and Monti Guarnieri, A. (1997), Multibaseline SAR
Interferometry for Automatic DEM Reconstruction, in: Third ERS Symposium—Space
at the Service of our Environment, Florence, Italy, 17–21 March 1997, ESA SP-414, pp.
1809–1820.
Feynman, R. P., Leighton, R. B. and Sands, M. (1963), The Feynman Lectures on Physics:
Commemorative Issue, vol. 1, Addison-Wesley company, Inc., Reading, Massachusetts.
Fialko, Y. and Simons, M. (2000), Deformation and seismicity in the Coso geothermal area,
Inyo County, California: Observations and modeling using satellite radar interferometry,
Journal of Geophysical Research, 105(B9):21781–21794.
Fielding, E., Wright, T., Parsons, B., England, P., Rosen, P., Hensley, S. and Bilham, R.
(1999), Topography of Northwest Turkey from SAR Interferometry: Applications to the
1999 Izmit earthquake Geomorphology and Co-seismic Strain, EOS Transactions, AGU,
80(46):F663.
Fitch, J. P. (1988), Synthetic aperture radar, Springer-Verlag, New York.
Flynn, T. J. (1997), Two-dimensional phase unwrapping with minimum weighted discon-
tinuity, Journal of the Optical Society of America A., 14:2692–2701.
Fornaro, G., Franceschetti, G. and Lanari, R. (1996a), Interferometric SAR Phase Unwrap-
ping Using Green’s Formulation, IEEE Transactions on Geoscience and Remote Sensing,
34(3):720–727.
Fornaro, G., Franceschetti, G., Lanari, R. and Sansosti, E. (1996b). Robust phase un-
wrapping Techniques: A Comparison, Journal of the Optical Society of America A.,
13(12):2355–2366.
Foster, M. R. and Guinzy, N. J. (1967), The coefficient of coherence: its estimation and
use in geophysical data processing, Geophysics, 32(4):602–616.
Franceschetti, G., Oliver, C. J., Shiue, J. C. and Tajbakhsh, S., eds. (1996), Microwave
Sensing and Synthetic Aperture Radar, SPIE, Bellingham.
Fruneau, B., Rudant, J.-P., Obert, D. and Raymond, D. (1998), Small displacements de-
tected by SAR interferometry on the city of Paris, (France), in: Second Int. Workshop on
”Retrieval of Bio- and Geophysical Parameters from SAR Data for Land Applications”
21-23 Oct, Noordwijk, The Netherlands, ESTEC.
Fujii, N., T Nakano, T. O. and Yamaoka, K. (1994), Detection of Ground Deformations by
the Multi–pass Differential SAR–interferometry. Examples of the Active Volcanic Area,
in: paper presented at the 1st workshop on SAR interferometry, Tokyo, Japan, December
 Bibliography 283

1994, NASDA.
Fujiwara, S., Yarai, H., Ozawa, S., Tobita, M., Murakami, M., Nakagawa, H., Nitta, K.,
Rosen, P. A. and Werner, C. L. (1998), Surface displacement of the March 26, 1997
Kagoshima-kenhokuseibu earthquake in Japan from synthetic aperture radar interfer-
ometry, Geophysical Research Letters, 25(24):4541–4544.
Fuliwara, S., Rosen, P., Tobita, M. and Murakami, M. (1998), Crustal deformation mea-
surments using repeat-pass JERS-1 synthetic aperture radar interferometry near the Izu
Peninsula, Japan, Journal of Geophysical Research, 103(B2):2411–2426.
Gabriel, A. K. and Goldstein, R. M. (1988), Crossed orbit interferometry: theory and
experimental results from SIR-B, Int.J. Remote Sensing, 9(5):857–872.
Gabriel, A. K., Goldstein, R. M. and Zebker, H. A. (1989), Mapping small elevation
changes over large areas: differential radar interferometry, Journal of Geophysical Re-
search, 94(B7):9183–9191.
Galloway, D. L., Hudnut, K. W., Ingebritsen, S. E., Phillips, S. P., Peltzer, G., Rogez, F. and
Rosen, P. A. (1998), Detection of acquifer system compaction and land subsidence using
interferometric synthetic aperture radar, Antelope Valley, Mojave Desert, California,
Water Resources Research, 34(10):2573–2585.
Gatelli, F., Monti Guarnieri, A., Parizzi, F., Pasquali, P., Prati, C. and Rocca, F. (1994),
The wavenumber shift in SAR Interferometry, IEEE Transactions on Geoscience and
Remote Sensing, 32(4):855–865.
Gauss, C. F. (1809), Theoria Motus Corporum Coelestium, Perthes und Besser, Hamburg.
Geudtner, D. (1995), Die Interferometrische Verarbeitung von SAR-Daten des ERS-1,
Ph.D. thesis, Universität Stuttgart, Draft.
Geudtner, D., Schwäbisch, M. and Winter, R. (1994), SAR-Interferometry with ERS-1
data, in: PIERS’94 ESA/ESTEC Noordwijk.
Geudtner, D., Winter, R. and Vachon, P. W. (1996), Flood mointoring using ERS-1 SAR
Interferometry coherence maps, Presented at IGARSS’96.
Ghiglia, D. C. and Pritt, M. D. (1998), Two-dimensional phase unwrapping: theory, algo-
rithms, and software, John Wiley & Sons, Inc, New York.
Ghiglia, D. C. and Romero, L. A. (1994), Robust two-dimensional weighted and unweighted
phase unwrapping that uses fast transforms and iterative methods, Journal of the Optical
Society of America A., 11(1):107–117.
Ghiglia, D. C. and Romero, L. A. (1996), Minimum two-dimensional phase un-
wrapping, Journal of the Optical Society of America A., 13:1999–2013.
Goldstein, R. (1995), Atmospheric limitations to repeat-track radar interferometry, Geo-
physical Research. Letters, 22(18):2517–2520.
Goldstein, R. M. (1964), Venus characteristics by earth-based radar, Astronomical Journal,
69(1):12–18.
Goldstein, R. M. (1969), Preliminary Venus Radar Results, Radio Science, 96D:1623–1625.
Goldstein, R. M., Caro, E. R. and Wu, C. (1985), Method and apparatus for contour
mapping using synthetic aperture radar, United States Patent, No. 4551724.
Goldstein, R. M. and Carpenter, R. L. (1963), Rotation of Venus: Period Estimated from
Radar Measurements, Science, 139:910–911.
Goldstein, R. M., Engelhardt, H., Kamp, B. and Frolich, R. M. (1993), Satellite radar
interferometry for monitoring ice sheet motion: Application to an Antarctic ice stream,
Science, 262:1525–1530.
Goldstein, R. M. and Gillmore, W. F. (1963), Radar observations of Mars, Science,
141:1171–1172.
Goldstein, R. M. and Werner, C. L. (1998), Radar interferogram filtering for geophysical
applications, Geophysical Research Letters, 25(21):4035–4038.
284 Bibliography

Goldstein, R. M. and Zebker, H. A. (1987), Interferometric radar measurement of ocean

surface currents, Nature, 328(20) :707–709.
Goldstein, R. M., Zebker, H. A. and Werner, C. L. (1988), Satellite radar interferometry:
Two-dimensional phase unwrapping, Radio Science, 23(4):713–720.
Golub, G. H. and Van Loan, C. F. (1996), Matrix Computations, Johns Hopkins University
Press, Baltimore, 3rd edn.
Goodman, J. W. (1975), Statistical Properties of Laser Speckle Patterns, in: Dainty (1975),
chap. 2, pp. 9–75.
Goodman, N. R. (1963), Statistical analysis based on a certain multivariate complex Gaus-
sian distribution (an introduction), Annals of Mathematical Statistics, 34(1):152–177.
Goovaerts, P. (1997), Geostatistics fo Natural Resources Evaluation, Applied Geostatistics
Series, Oxford University Press, New York.
Gradshteyn, I. S., Ryzhik, I. M. and Jeffrey, A. (1994), Table of integrals, series, and
products, Academic Press, Boston, 5th edn., Original Title: Tablitsy integralov, summ,
ryadov i proizvediniy, 1965.
Graham, L. C. (1974), Synthetic Interferometer Radar for Topographic Mapping, Proceed-
ings of the IEEE, 62(6):763–768.
Gray, A. L. and Farris-Manning, P. J. (1993), Repeat-Pass Interferometry with Airborne
Synthetic Aperture Radar, IEEE Transactions on Geoscience and Remote Sensing,
31(1):180–191.
Gray, A. L., Mattar, K. E. and Sofko, G. (2000), Influence of Ionospheric Electron
Density Fluctuations on Satellite Radar Interferometry, Geophysical Research Letters,
27(10):1451–1454.
Gray, L., Mattar, K. and Short, N. (1999), Speckle Tracking for 2-Dimensional Ice Motion
Studies in Polar Regions: Influence of the Ionosphere, in: Second International Workshop
on ERS SAR Interferometry, ‘FRINGE99’, Liège, Belgium, 10–12 Nov 1999, pp. 1–8,
ESA.
Hagberg, J. O., Ulander, L. M. H. and Askne, J. (1995), Repeat-Pass SAR Interferom-
etry over Forested Terrain, IEEE Transactions on Geoscience and Remote Sensing,
33(2):331–340.
Hall, M. P. M., Barclay, L. W. and Hewitt, M. T., eds. (1996), Propagation of Radiowaves,
The Institution of Electrical Engineers, London.
Haltiner, G. J. and Martin, F. L. (1957), Dynamical and Physical Meteorology, McGraw-
Hill, New York.
Hannan, E. J. and Thomson, P. J. (1971), The estimation of coherence and group delay,
Biometrika, 58(3-481):469.
Hanssen, R. (1994), The Application of Radon transformation for improved analysis of
sparsely sampled data: a feasibility study, Faculty of Geodetic Engineering, Rep.93.1,
Delft University of Technology, Delft.
Hanssen, R. (1996), GISARE: GPS metingen voor de schatting van atmosferische param-
eters, GPS Nieuwsbrief, 11(1):17–21, In Dutch.
Hanssen, R. (1998), Atmospheric heterogeneities in ERS tandem SAR interferometry, Delft
University Press, Delft, the Netherlands.
Hanssen, R., Amelung, F. and Zebker, H. (1998a), Geodetic interpretation of land subsi-
dence measurements at the Cerro Prieto geothermal field monitored by radar interfer-
ometry, EOS Transactions, AGU, 79(45):F37.
Hanssen, R. and Bamler, R. (1999), Evaluation of Interpolation Kernels for SAR Interfer-
ometry, IEEE Transactions on Geoscience and Remote Sensing, 37(1):318–321.
Hanssen, R. and Feijt, A. (1996), A first quantitative evaluation of atmospheric effects on
SAR interferometry, in: ‘FRINGE 96’ workshop on ERS SAR Interferometry, Zürich,
 Bibliography 285

Switzerland, 30 Sep–2 October 1996, pp. 277–282, ESA SP-406.

Hanssen, R. and Usai, S. (1997), Interferometric phase analysis for monitoring slow defor-
mation processes, in: Third ERS Symposium—Space at the Service of our Environment,
Florence, Italy, 17–21 March 1997, ESA SP-414, pp. 487–491.
Hanssen, R., Vermeersen, B., Scharroo, R., Kampes, B., Usai, S., Gens, R. and Klees, R.
(2000a), Deformatiepatroon van de aardbeving van 17 augustus 1999 in Turkije gemeten
met satelliet radar interferometrie, Remote Sensing Nieuwsbrief, 90:42–44, In Dutch.
Hanssen, R., Weinreich, I., Lehner, S. and Stoffelen, A. (2000b), Tropospheric wind and
humidity derived from spaceborne radar intensity and phase observations, Geophysical
Research Letters, 27(12):1699–1702.
Hanssen, R., Zebker, H., Klees, R. and Barlag, S. (1998b), On the use of meteorological
observations in SAR interferometry, in: International Geoscience and Remote Sensing
Symposium, Seattle, Washington, USA, 6–10 July 1998, pp. 1644–1646.
Hanssen, R. F., Feijt, A. J. and Klees, R. (2001), Comparison of water vapor column derived
from radar interferometry and Meteosat WV channel analysis, Journal of Atmospheric
and Oceanic Technology, in press.
Hanssen, R. F. and Weckwerth, T. M. (1999), Evaluation of spatial moisture distribution
during CLARA 96 using spaceborne radar interferometry, in: Remote Sensing of cloud
parameters: retrieval and validation. 21-22 Oct 1999, Delft, The Netherlands, pp. 15–20.
Hanssen, R. F., Weckwerth, T. M., Zebker, H. A. and Klees, R. (1999), High-Resolution
Water Vapor Mapping from Interferometric Radar Measurements, Science, 283:1295–
1297.
Hartl, P. (1991), Application of Interferometric SAR-Data of the ERS-1 Mission for High
Resolution Topographic Terrain Mapping, GIS, 2:8–14.
Hartl, P., Reich, M., Thiel, K.-H. and Xia, Y. (1993), SAR interferometry applying ERS-1:
some preliminary test results, in: First ERS-1 Symposium—Space at the Service of our
Environment, Cannes, France, 4–6 November 1992, ESA SP-359, pp. 219–222.
Hartl, P. and Thiel, K.-H. (1993), Bestimmung von topographischen Feinstrukturen mit
interferometrischem ERS-1-SAR, Zeitschrift für Photogrammetrie und Fernerkundung,
3:108–114.
Hartl, P., Thiel, K.-H. and Wu, X. (1994a), Information extraction from ERS-1 SAR data
by means of INSAR and D-INSAR techniques in Antarctic research, in: Second ERS-
1 Symposium—Space at the Service of our Environment, Hamburg, Germany, 11–14
October 1993, ESA SP-361, pp. 697–701.
Hartl, P., Thiel, K. H., Wu, X., Doake, C. and Sievers, J. (1994b), Application of SAR
Interferometry with ERS-1 in the Antarctic, Earth Observation Quarterly, 43:1–4.
Hartl, P. and Xia, Y. (1993), Besonderheiten der Datenverarbeitung bei der SAR-
Interferometrie, Zeitschrift für Photogrammetrie und Fernerkundung, 6:214–222.
Henderson, F. M. and Lewis, A. J., eds. (1998), Principles and applications of Imaging
Radar, vol. 2 of Manual of Remote Sensing, John Wiley & Sons, Inc., New York, 3rd
edn.
Hernandez, B., Cotton, F., Campillo, M. and Massonnet, D. (1997), A Comparison between
short term (co-seismic) and long term (one year) slip for the Landers earthquake: Mea-
surements from strong motion and SAR interferometry, Geophysical Research Letters,
24(13):1579–1582.
Herring, T. A. (1992), Modeling atmospheric delays in the analysis of space geodetic data,
in: de Munck, J. C. and Spoelstra, T. A. T., eds., Symposium on Refraction of trans at-
mospheric signals in geodesy, pp. 157–164, Delft, Netherlands Geodetic Commission.
Ho, C. M., Mannucci, A. J., Lindqwister, U. J., Pi, X., Tsurutani, B. T., Sparks, L.,
Iijima, B. A., Wilson, B. D., Harris, I. and Reyes, M. J. (1998), Global ionospheric TEC
286 Bibliography

variations during January 10, 1997 storm, Geophysical Research Letters, 25(14):2589–
2592.
Hoen, E. W. and Zebker, H. A. (2000), Penetration Depths Inferred from Interferometric
Volume Decorrelation Observed over the Greenland Ice Sheet, IEEE Transactions on
Geoscience and Remote Sensing, 38(6):2571–2583.
van der Hoeven, A. G. A., Ambrosius, B. A. C., van der Marel, H., Derks, H., Klein
Baltink, H., van Lammeren, A. and Kösters, A. J. M. (1998), Analysis and Comparison
of Integrated Water Vapor Estimation from GPS, in: Proc. llth Intern. Techn. Meeting
Sat. Div. Instit. of Navigation, pp. 749–755.
van der Hoeven, A. G. A. and Hanssen, R. F. (1999), Cross-Validation of Tropospheric
Water Vapor Measurements by GPS and SAR interferometry, in: General Assembly
European Geophysical Society, The Hague, The Netherlands, 19-23 April 1999.
Hogg, D. C., Guiraud, F. O. and Dekker, M. T. (1981), Measurement of Excess Radio
Transmission Length on Earth-space Paths, Astronomy and Astrophysics, 95:304–307.
Homer, J. and Longstaff, I. D. (1995), Improving Height Resolution of Interferometric
SAR, in: International Geoscience and Remote Sensing Symposium, Florence, Italy,
10–14 July 1995, pp. 210–212.
Hovanessian, S. A. (1980), Introduction to Synthetic Array and Imaging Radars, Artech
House, Inc., Dedham, MA.
Hubert-Ferrari, A., Barka, A., Jacques, E., Nalbant, S. S., Meyer, B., Armijo, R., Tappon-
nier, P. and King, G. C. P. (2000), Seismic hazard in the Marmara Sea region following
the 17 August 1999 Izmit earthquake, Nature, 404:269–273.
Hulsmeyer, C. (1904), Herzian-Wave Projecting and Receiving Apparatus Adopted to In-
dicate or Give Warning of the Presence of a Metallic Body such as a Ship or a Train,
British Patent No. 13,170.
Hunt, B. R. (1979), Matrix formulation of the reconstruction of phase values from phase
differences, Journal of the Optical Society of America A., 69(3):393–399.
Huschke, R. E. (1959), Glossary of meteorology, American Meteorological Society, Boston.
Huygens, C. (1690), Traité de la lumière, où sont expliquées les causes de ce qui luy arrive
dans la reflexion, et dans la refraction: et particulierement dans l’étrange refraction du
cristal d’Islande: avec un discours de la cause de la pesanteur, Van der Aa, Leiden.
Ishimaru, A. (1978), Wave Propagation and Scattering in Random Media, vol. 2, Academic
Press, New York.
Jakowski, N., Bettac, H. D. and Jungstand, A. (1992), Ionospheric corrections for Radar
Altimetry and Geodetic Positioning Techniques, in: Proc. of the Symp. Refraction of
Trans atmospheric Signals in Geodesy, the Hague, 19–22 may 1992, pp. 151–154.
Jónsson, S., Adam, N. and Björnsson, H. (1998), Effects of Subglacial Geothermal Activity
Observed by Satellite Radar Interferometry, Geophysical Research Letters, 25(7): 1059–
1062.
Jónsson, S., Zebker, H., Cervelli, P., Segall, P., Garbeil, H., Mouginis-Mark, P. and Row-
land, S. (1999), A Shallow-Dipping Dike fed the 1995 Flank Eruption at Fernandina
Volcano, Galapagos, Observed by Satellite Radar Interferometry, Geophysical Research
Letters, 26(8): 1077–1080.
Jordan, R. L., Caro, E. R., Kim, Y., Kobrick, M., Shen, Y., Stuhr, F. V. and Werner,
M. U. (1996), Shuttle radar topography mapper (SRTM), in: Franceschetti et al. (1996),
pp. 412–422.
Joughin, L, Gray, L., Bindschadler, R., Price, S., Morse, D., Hulbe, C., Mattar, K. and
Werner, C. (1999), Tributaries of West Antarctic Ice Streams Revealed by RADARSAT
Interferometry, Science, 286(5438):283–286.
Joughin, I. R. (1995), Estimation of Ice-Sheet Topography and Motion Using Interferometric
 Bibliography 287

Synthetic Aperture Radar, Ph.D. thesis, University of Washington.

Joughin, I. R., Fahnestock, M., Ekholm, S. and Kwok, R. (1997), Balance Velocities of the
Greenland ice sheet, Geophysical Research Letters, 24(23) :3045–3048.
Joughin, I. R., Kwok, R. and Fahnestock, M. A. (1998), Interferometric Estimation of
Three-Dimensional Ice-Flow Using Ascending and Descending Passes, IEEE Transac-
tions on Geoscience and Remote Sensing, 36(1):25–37.
Joughin, I. R., Winebrenner, D., Fahnestock, M., Kwok, R. and Krabill, W. (1996), Mea-
surement of ice-sheet topography using satellite-radar interferometry, Journal of Glaciol-
ogy, 42(140):10–22.
Joughin, I. R. and Winebrenner, D. P. (1994), Effective Number of Looks for a Multilook
Interferometric Phase Distribution, in: International Geoscience and Remote Sensing
Symposium, Pasadena, CA, USA, 8–12 August 1994, pp. 2276–2278.
Journel, A. G. and Huijbregts, C. J. (1978), Mining Geostatistics, Academic Press, London.
Just, D. and Bamler, R. (1994), Phase statistics of interferograms with applications to
synthetic aperture radar, Applied Optics, 33(20):4361–4368.
Kailath, T., Vieira, A. and Morf, M. (1978), Inverses of Toeplitz Operators, innovations,
and orthogonal polynomials, SIAM Review, 20(1):106–119.
Kawai, S. and Shimada, M. (1994), Detections of Earth Surface Deformation Change by
means of INSAR technique, in: paper presented at the 1st workshop on SAR interferom-
etry, Tokyo, Japan, December 1994, NASDA.
Kellogg, O. D. (1929), Foundations of Potential Theory, Frederick Ungar, New York.
Keys, R. G. (1981), Cubic Convolution Interpolation for Digital Image Processing, IEEE
Transactions on Acoustics, Speech, and Signal Processing, ASSP-29(6):1153–1160.
King, G. C. P., Stein, R. S. and Lin, J. (1994), Static stress changes and the triggering of
earthquakes, Bulletin of the Seismological Society of America, 84:935–953.
Klinger, Y., Michel, R. and Avouac, J.-P. (2000), Co-seismic deformation during the Mw 7.3
Aqaba earthquake (1995) from ERS-SAR interferometry, Geophysical Research Letters,
27(22):3651–3655.
Koch, K.-R. (1999), Parameter Estimation and Hypothesis Testing in Linear Models,
Springer-Verlag, New York, 2nd edn.
Kolmogorov, A. N. (1941), Dissipation of energy in locally isotropic turbulence, Doklady
Akad. Nauk SSSR, 32(16), German translation in “Sammelband zur Statistischen The-
orie der Turbulenz”, Akademie Verlag, Berlin, 1958, p77.
van der Kooij, M. (1999a), Personal communication.
van der Kooij, M. (1999b), Production of DEMs from ERS tandem Data, in: Second
International Workshop on ERS SAR Interferometry, ‘FRINGE99’, Liège, Belgium, 10–
12 Nov 1999, ESA.
van der Kooij, M. W. A., van Halsema, D., Groenewoud, W., Mets, G. J., Overgaauw, B.
and Visser, P. N. A. M. (1995), SAR Land Subsidence Monitoring, vol. NRSP-2 95-13,
Beleidscommissie Remote Sensing, Delft.
Krul, L. (1982), Principles of radar measurement, in: Radar calibration, Proceedings of
EARSeL Workshop, Alpbach, Austria, 6–10 December 1982, pp. 11–20.
Kuettner, J. P. (1971), Cloud bands in the earth’s atmosphere, Tellus, 23:404–425.
Kursinski, E. R. (1997), The GPS radio occultation concept: theoretical performance and
initital results, Ph.D. thesis, California Institute of Technology.
Kursinski, E. R., Haij, G. A., Schofield, J. T. and Linfield, R. P. (1997), Observing Earth’s
atmosphere with radio occultation measurements using the Global Positioning System,
Journal of Geophysical Research, 102(D19):23,429–23,465.
Kwok, R. and Fahnestock, M. A. (1996), Ice Sheet Motion and Topography from Radar
Interferometry, IEEE Transactions on Geoscience and Remote Sensing, 34(1): 189–200.
288 Bibliography

van Lammeren, A. C. A. P., Russchenberg, H. W. J., Apituley, A. and ten Brink, H. (1997),
CLARA: a data set to study sensor synergy, in: Proceedings Workshop on Synergy
of Active Instruments in the Earth Radiation Mission, November 12-14, Geesthacht,
Germany, ESA EWP 1968 or GKSS 98/eE10.
Lanari, R., Fornaro, G., Riccio, D., Migliaccio, M., Papathanassiou, K. P., ao R Moreira, J.,
Schwäbisch, M., Dutra, L., Puglisi, G., Franceschetti, G. and Coltelli, M. (1996), Gen-
eration of Digital Elevation Models by Using SIR-C/X-SAR Multifrequency Two-Pass
Interferometry: The Etna Case Study, IEEE Transactions on Geoscience and Remote
Sensing, 34(5):1097–1114.
Lanari, R., Lundgren, P. and Sansosti, E. (1998), Dynamic deformation of Etna volcano
observed by satellite radar interferometry, Geophysical Research Letters, 25:1541–1544.
Lay, O. P. (1997), The temporal power spectrum of atmospheric fluctuations due to water
vapor, Astronomy and Astrophysics Supplement Series, 122:535–545.
Lee, J. S., Hoppel, K. W., Mango, S. A. and Miller, A. R. (1994), Intensity and phase
statistics of multilook polarimetric and interferometric SAR imagery, IEEE Transactions
on Geoscience and Remote Sensing, 30:1017.
Lehner, S., Horstmann, J., Koch, W. and Rosenthal, W. (1998), Mesoscale wind mea-
surements using recalibrated ERS SAR images, Journal of Geophysical Research,
103(C4):7847–7856.
LeMone, M. A. (1973), The structure and dynamics of horizontal roll vortices in the plan-
etary boundary layer, J. Atmos. Sci., 30:1077–1091.
LeMone, M. A. and Pennell, W. T. (1976), The relationship of trade wind cumulus distri-
bution to subcloudlayer fluxes and structure, Monthly Weather Review, 104:524–539.
Li, F. and Goldstein, R. (1987), Studies of Multi-baseline Spaceborne Interferometric Syn-
thetic Aperture Radar, in: International Geoscience and Remote Sensing Symposium,
Ann Arbor, 18–21 May 1987, pp. 1545–1550.
Li, F. K. and Goldstein, R. M. (1990), Studies of Multibaseline Spaceborne Interferomet-
ric Synthetic Aperture Radars, IEEE Transactions on Geoscience and Remote Sensing,
28(l):88–97.
Lin, Q., Vesecky, J. F. and Zebker, H. A. (1992), New approaches in Interferometric SAR
data processing, IEEE Transactions on Geoscience and Remote Sensing, 30(3):560–567.
Lu, Z., Fatland, R., Wyss, M., Li, S., Eichelberger, J., Dean, K. and Freymueller, J. (1997),
Deformation of New Trident Volcano measured by ERS 1 SAR interferometry, Katmai
National Park, Alaska, Geophysical Research Letters, 24(6):695–698.
Lu, Z. and Freymueller, J. T. (1999), Synthetic aperture radar interferometry coher-
ence analysis over Katmai volcano group, Alaska, Journal of Geophysical Research,
103(B12):29887–29894.
Madsen, S. N. (1986), Speckle Theory: Modelling, analysis, and applications related to
Synthetic Aperture Radar Data, Ph.D. thesis, Technical University of Denmark.
Madsen, S. N. (1989), Estimating the Doppler Centroid of SAR Data, IEEE Transactions
on Aerospace and Electronic Systems, 25(2): 134–140.
Mandelbrot, B. B. (1983), The Fractal Geometry of Nature, Freeman, New York.
Mandelbrot, B. B. and Ness, J. W. V. (1968), Fractional Brownian motions, fractional
noises and applications, SIAM Review, 10(4):422–438.
Margot, J. L., Campbell, D. B., Jurgens, R. F. and Slade, M. A. (1999a), Topography of
the Lunar Poles from Radar Interferometry: A Survey of Cold Trap Locations, Science,
284:1658–1660.
Margot, J.-L., Campbell, D. B., Jurgens, R. F. and Slade, M. A. (1999b), The topography
of Tycho Crater, Journal of Geophysical Research, 104(E5):11875–11882.
Markoff, A. A. (1912), Wahrscheinlichkeitsrechnung, Teubner, Leibzig.
 Bibliography 289

Massonnet, D. and Adragna, F. (1993), A full-scale validation of Radar Interferometry with

ERS-1: the Landers earthquake, Earth Observation Quarterly, 41.
Massonnet, D., Adragna, F. and Rossi, M. (1994a), CNES General-Purpose SAR Correla-
tor, IEEE Transactions on Geoscience and Remote Sensing, 32(3):636–643.
Massonnet, D., Briole, P. and Arnaud, A. (1995), Deflation of Mount Etna monitored by
spaceborne radar interferometry, Nature, 375:567–570.
Massonnet, D., Feigl, K., Rossi, M. and Adragna, F. (1994b), Radar interferometric map-
ping of deformation in the year after the Landers earthquake, Nature, 369:227–230.
Massonnet, D. and Feigl, K. L. (1995a), Discrimination of geophysical phenomena in satel-
lite radar interferograms, Geophysical Research Letters, 22(12): 1537–1540.
Massonnet, D. and Feigl, K. L. (1995b), Satellite radar interferometric map of the coseismic
deformation field of the M = 6.1 Eureka Valley, California earthquake of May 17, 1993,
Geophysical Research Letters, 22(12):1541–1544.
Massonnet, D. and Feigl, K. L. (1998), Radar interferometry and its application to changes
in the earth’s surface, Reviews of Geophysics, 36(4):441–500.
Massonnet, D., Feigl, K. L., Vadon, H. and Rossi, M. (1996a), Coseismic deformation field
of the M=6.7 Northridge, California earthquake of January 17, 1994 recorded by two
radar satellites using interferometry, Geophysical Research Letters, 23(9):969–972.
Massonnet, D., Holzer, T. and Vadon, H. (1997), Land subsidence caused by the East Mesa
geothermal field, California, observed using SAR interferometry, Geophysical Research
Letters, 24(8):901–904, Correction in GRL v.25,n.16,p.3213,1998.
Massonnet, D., Rossi, M., Carmona, C., Adagna, F., Peltzer, G., Feigl, K. and Rabaute,
T. (1993), The displacement field of the Landers earthquake mapped by radar interfer-
ometry, Nature, 364(8):138–142.
Massonnet, D., Thatcher, W. and Vadon, H. (1996b), Detection of postseismic fault zone
collapse following the Landers earthquake, Nature, 382:489–497.
Massonnet, D. and Vadon, H. (1995), ERS-1 Internal Clock Drift Measured by Interferom-
etry, IEEE Transactions on Geoscience and Remote Sensing, 33(2):401–408.
Massonnet, D., Vadon, H. and Rossi, M. (1996c), Reduction of the Need for Phase Unwrap-
ping in Radar Interferometry, IEEE Transactions on Geoscience and Remote Sensing,
34(2):489–497.
McCarthy, D. D., ed. (1996), IERS Conventions (1996), IERS Technical Note 21, Obser-
vatoire de Paris.
Melsheimer, C., Alpers, W. and Gade, M. (1996), Investigation of multifre-
quency/multipolarization radar signatures of rain cells over the ocean using SIR-C/X-
SAR data, submitted to JGR.
Meyer, B., Armijo, R., Massonnet, D., de Chabalier, J. B., Delacourt, C., Ruegg, J. C.,
Achache, J., Briole, P. and Panastassiou, D. (1996), The 1995 Grevena (Northern Greece)
earthquake: fault model constrained with tectonic observations and SAR interferometry,
Geophysical Research Letters, 23:2677–2680.
Michel, R., Avouac, J.-P. and Taboury, J. (1999), Measuring ground displacement from
SAR amplitude images: application to the Landers earthquake, Geophysical Research
Letters, 26(7):875–878.
Mogi, K. (1958), Relations Between Eruptions of Various Volcanoes and the Deformation
of the Ground Surface Around Them, Bulletin of the Earthquake Research Institute,
University of Tokyo, 36:99–134.
Mohr, J. J. (1997), Repeat Track SAR Interferometry. An investigation of its Utility for
Studies of Glacier Dynamics, Ph.D. thesis, Technical University of Denmark, Copen-
hagen.
Mohr, J. J., Reeh, N. and Madsen, S. N. (1998), Three-dimensional glacial flow and surface
290 Bibliography

elevation measured with radar interferometry, Nature, 291:273–276.

Monin, A. S., Yaglom, A. M. and Lumley, J. L. (1975), Statistical Fluid Mechanics: Me-
chanics of Turbulence, vol. 2, MIT Press, Cambridge, Original title: Statisicheskaya
gidromekhanika – Mekhanika Turbulenosti.
Monti Guarnieri, A. and Prati, C. (1997), SAR Interferometry: A “Quick and Dirty”
Coherence Estimator for Data Browsing, IEEE Transactions on Geoscience and Remote
Sensing, 35(3):660–669.
Mossop, A., Murray, M., Owen, S. and Segall, P. (1997), Subsidence at the Geysers geother-
mal field: results and simple models, in: Twenty-Second Workshop on Geothermal Reser-
voir Engineering, pp. 377–382, Stanford University, Jan 27-29, 1997.
Mouginis-Mark, P. J. (1995a), Analysis of volcanic hazards using Radar interferometry,
Earth Observation Quarterly, 47:6–10.
Mouginis-Mark, P. J. (1995b). Preliminary Observations of Volcanoes with the SIR-C
Radar, IEEE Transactions on Geoscience and Remote Sensing, 33(4):934–941.
Mourad, P. D. and Walter, B. A. (1996), Viewing a cold air outbreak using satellite-based
synthetic aperture radar and advanced very high resolution radiometer imagery, Journal
of Geophysical Research, 101(C7):16391–16400.
Mueller, C. K., Wilson, J. W. and Crook, N. A. (1993), The utility of sounding and mesonet
data to nowcast thunderstorm initiation, Weather and Forecasting, 8:132–146.
Murakami, M., Tobita, M., Fujiwara, S. and Saito, T. (1996), Coseismic crustal defor-
mations of 1994 Northridge, California, earthquake detected by interferometric JERS-1
Synthetic aperture radar, Journal of Geophysical Research, 101(B4):8605–8614.
NAM (2000), Bodemdaling door Aardgaswinning: Groningen veld en randvelden in Gronin-
gen, Noord-Drenthe en het Oosten van Friesland. Status Rapport 2000 en Prognose tot
het jaar 2050., NAM rapport nr. 2000-02-00410, Nederlandse Aardolie Maatschappij
B.V, Assen, in Dutch.
Newland, D. E. (1993), Random vibrations, spectral and wavelet analysis, Longman, Harlow,
3rd edn.
Niell, A. E. (1996), Global mapping functions for the atmosphere delay at radio wave-
lengths, Journal of Geophysical Research, 101(B2):3227–3246.
NorthWest Research Associates, I. (2000), Atmospheric Sciences, (Accessed 15 Jun, 2000)
<http://www.nwra.com/>.
Oberhettinger. F. (1970), Hypergeometric Functions, in: Abramowitz and Stegun (1970),
chap. 15, pp. 555–566.
Obukov, A. M. (1941), On the distribution of energy in the spectrum of turbulent flow,
Doklady Akad. Nauk SSSR, 32(19).
Ohkura, H. (1998), Applications of SAR data to monitoring Earth surface changes and
displacements, Advances in Space Research, 21(3):485–492.
Olmsted, C. (1993), Alaska SAR Facility Scientific SAR User’s Guide.
Otten, M. P. G. (1998), Review of SAR processing techniques, in: 28th EuMC workshop
proceedings, pp. 33–41, EuMW Amsterdam, 9 October 1998.
Ozawa, S., Murakami, M., Fujiwara, S. and Tobita, M. (1997), Synthetic aperture radar
interferogram of the 1995 Kobe earthquake and its geodetic inversion, Geophysical Re-
search Letters, 24(18):2327–2330.
Papoulis. A. (1968), Systems and Tronsforms with Applications in Optics, McGraw-Hill
series in Systems Science, McGraw-Hill, New York.
Papoulis, A. (1991), Probability, Random variables, and stochastic processes, McGraw-Hill
series in Electrical Engineering, McGraw-Hill, New York.
Park, S. K. and Schowengerdt, R. A. (1983), Image Reconstruction by Parametric Cubic
Convolution, Computer Vision, Graphics and Image Processing, 23(3):258–272.
 Bibliography 291

Parsons, D. B. (1992), An Explanation for Intense Frontal Updrafts and Narrow Cold-
Frontal Rainbands, Journal of the Atmospheric Sciences, 49(19):1810–1825.
Parsons, T., Toda, S., Stein, R. S., Barka, A. and Dieterich, J. H. (2000), Heightened
Odds of Large Earthquakes Near Istanbul: An Interaction-Based Probability Calculation,
Science, 288:661–665.
Peitgen, H. O. and Saupe, D., eds. (1988), The Science of Fractal Images, Springer-Verlag,
New York.
Peltzer, G., Crampé, F. and King, G. (1999), Evidence of the Nonlinear Elasticity of the
Crust from Mw7.6 Manyi (Tibet) Earthquake, Science, 286(5438):272–276.
Peltzer, G., Hudnut, K. W. and Feigl, K. L. (1994), Analysis of coseismic surface displace-
ment gradients using radar interferometry: New insights into the Landers earthquake,
Journal of Geophysical Research, 99(B11):21971–21981.
Peltzer, G. and Rosen, P. (1995), Surface Displacement of the 17 May 1993 Eureka Valley,
California Earthquake Observed by SAR Interferometry, Science, 268:1333–1336.
Peltzer, G., Rosen, P., Rogez, F. and Hudnut, K. (1996), Postseismic rebound in fault
step-overs caused by pore fluid flow, Science, 273:1202–1204.
Prati, C. and Rocca, F. (1990), Limits to the resolution of elevation maps from stereo SAR
images, Int. J. Remote Sensing, 11(12):2215–2235.
Prati, C. and Rocca, F. (1992), Range Resolution Enhancement with multiple SAR Surveys
Combination, in: International Geoscience and Remote Sensing Symposium, Houston,
Texas, USA, May 26–29 1992, pp. 1576–1578.
Prati, C. and Rocca, F. (1993), Improving Slant-Range Resolution With Multiple SAR
Surveys, IEEE Transactions on Aerospace and Electronic Systems, 29(1): 135–144.
Prati, C. and Rocca, F. (1994a), DEM generation with ERS-1 interferometry, in: Sansò,
F., ed., Geodetic Theory Today, pp. 19–26, Berlin, Third Hotine-Marussi Symposium on
Mathematical Geodesy, Springer-Verlag.
Prati, C. and Rocca, F. (1994b), Use of the spectral shift in SAR interferometry, in: Second
ERS-1 Symposium—Space at the Service of our Environment, Hamburg, Germany, 11–
14 October 1993, ESA SP-361, pp. 691–696.
Prati, C., Rocca, F. and Monti Guarnieri, A. (1989), Effects of speckle and additive noise on
the altimetric resolution of interferometric SAR (ISAR) surveys, in: Proc. IGARSS’89,
Vancouver, pp. 2469–2472.
Prati, C., Rocca, F. and Monti Guarnieri, A. (1993), SAR interferometry experiments with
ERS-1, in: First ERS-1 Symposium—Space at the Service of our Environment, Cannes,
France, 4–6 November 1992, ESA SP-359, pp. 211–217.
Prati, C., Rocca, F., Monti Guarnieri, A. and Damonti, E. (1990), Seismic Migrataion For
SAR Focussing: Interferometrical Applications, IEEE Transactions on Geoscience and
Remote Sensing, 28(4):627–640.
Price, E. J. (1999), Coseismic and Postseismic Deformations Associated With the 1992
Landers, California Earthquake Measured by Synthetic Aperture Radar Interferometry,
Ph.D. thesis, University of California, San Diego.
Price, E. J. and Sandwell, D. T. (1998), Small-scale deformations associated with the
1992 Landers, California, earthquake mapped by synthetic aperture radar interferometry
phase gradients, Journal of Geophysical Research, 103(B11):27001-27016.
Priestley, M. B. (1981), Spectral Analysis and Time Series, Academic Press, London.
Pritt, M. D. (1996), Phase Unwrapping by Means of Multigrid Techniques for Interfero-
metric SAR, IEEE Transactions on Geoscience and Remote Sensing, 34(3):728–738.
Radarsat (1999), RADARSAT Specifications, (Accessed 2 Nov, 1999),
<http://radarsat.space.gc.ca/info/description/specifications.html>.
Raney, R. K. (1998), Radar Fundamentals: Technical Perspective, in: Henderson and Lewis
292 Bibliography

(1998), chap. 2, pp. 9–130.

Raney, R. K., Runge, H., Bamler, R., Cumming, I. G. and Wong, F. H. (1994), Precision
SAR Processing Using Chirp Scaling, IEEE Transactions on Geoscience and Remote
Sensing, 32(4):786–799.
Reilinger, R. E., Ergintav, S., Bürgmann, R., McClusky, S., Lenk, O., Barka, A., Gurkan,
O., Hearn, L., Feigl, K. L., Cakmak, R., Aktug, B., Ozener, H. and Töksoz, M. N. (2000),
Coseismic and Postseismic Fault Slip for the 17 August 1999, M = 7.5, Izmit, Turkey
Earthquake, Science, 289(5484):1519–1524.
Resch, G. M. (1984), Water Vapor Radiometry in Geodetic Applications, in: Brunner,
F. K., ed., Geodetic refraction, pp. 53–84, Springer-Verlag, Berlin.
Richelson, J. T. (1991), The future of Space Reconnaisance, Scientific American,
264(l):38–45.
Richman, D. (1982). Three-dimensional Azimuth-correcting Mapping Radar, United States
Patent, No. 4321601, Orgininally filed in 1971.
Rignot, E. J., Gogineni, S. P., Krabill, W. B. and Ekholm, S. (1997), North and northeast
Greenland ice discharges from satellite radar interferometry, Science, 276:934–937.
Rind, D., Chiou, E. W., Chu, W. P., Larsen, J. C., Oltmans, S., Lerner, J., McCormick,
M. P. and McMaser, L. R. (1991), Positive water vapor feedback in climate models
confirmed by satellite data, Nature, 349:7500–7503.
Riva, R., Hanssen, R., Vermeersen, L. L. A., Aoudia, K. and Sabadini, R. (2000), Coseismic
deformation of the 17 Aug 1999, Izmit, Turkey earthquake from normal mode modeling
and satellite radar interferometry (SAR) imaging, in: EGS XXV General Assembly,
Nice, France, 25-29 April 1999, Geophys. Res Abstracts, 2.
Rodriguez, E. and Martin, J. M. (1992), Theory and design of interferometric synthetic
aperture radars, IEE Proceedings-F, 139(2):147–159.
Rogers, A. E. and Ingalls, R. P. (1969), Venus: Mapping the surface reflectivity by Radar
Interferometry, Science, 165(3895):797–799.
Rosen, P., Werner, C., Fielding, E., Hensley, S. and Vincent, S. B. P. (1998), Aseismic
creep along the San Andreas fault northwest of Parkfield, CA measured by radar inter-
ferometry, Geophysical Research Letters, 25(6):825–828.
Rosen, P. A., Hensley, S., Zebker, H. A., Webb, F. H. and Fielding, E. J. (1996), Surface
deformation and coherence measurements of Kilauea Volcano, Hawaii, from SIR-C radar
interferometry, Journal of Geophysical Research, 101(E10):23109–23125.
Roth, A., Adam, N., Schwäbisch, M., Müschen, B., Böhm, C. and Lang, O. (1997), Ob-
servation of the effects of the subglacial volcano eruption underneath the Vatnajöküll
glacier in Iceland with ERS-SAR data, in: Proc. of the third ERS symposium, Florence,
Italy, 17-20 March 1997.
Rothacher, M. and Mervart, L., eds. (1996), Bernese GPS Software version 4.0, Astronom-
ical Institute University of Berne, Berne.
Rott, H., Schechl, B., Siegel, A. and Grasemann, B. (1999), Monitoring very slow slope
movements by means of SAR interferometry: A case study from a mass waste above a
reservoir in the Ötztal Alps, Austria, Geophysical Research Letters, 26(11):1629–1632.
Rott, H. and Siegel, A. (1997), Glaciological Studies in the Alps and in Antarctica Using
ERS Interferometric SAR, in: ‘FRINGE 96’ workshop on ERS SAR Interferometry,
Zürich, Switzerland, 30 Sep–2 October 1996, pp. 149–159, ESA SP-406, Vol II.
Rott, H., Stuefer, M., Siegel, A., Skvarca, P. and Eckstaller, A. (1998), Mass fluxes and
dynamics of Moreno Glacier, Southern Patagonia Icefield, Geophysical Research Letters,
25(9):1407–1410.
Ruf, C. S. and Beus, S. E. (1997), Retrieval of Tropospheric Water Vapor Scale Height
from Horizontal Turbulence Structure, IEEE Transactions on Geoscience and Remote
 Bibliography 293

Sensing, 35(2):203–211.
Rummel, R. (1990), Physical Geodesy 2, Lecture Notes, Delft University of Technology,
Department of Geodetic Engineering, Delft.
Runge, H. and Bamler, R. (1992), A novel high-precision SAR processing algorithm based
on chirp scaling, in: International Geoscience and Remote Sensing Symposium, Houston,
Texas, USA, May 26–29 1992.
Russchenberg, H. (2000), Personal communication.
Ryle, M. and Vonberg, D. D. (1946), Solar radiation on 175 Mc/s, Nature, 158:339–340.
Saastamoinen, J. (1972), Introduction to Practical Computation of Astronomical Refrac-
tion, Bulletin Geodesique, 106:383–397.
Saito, A., Fukao, S. and Miyazaki, S. (1998). High resolution mapping of TEC perturbations
with the GSI GPS network over Japan, Geophysical Research Letters, 25(16):3079–3082.
Samson, J. (1996), Coregistration in SAR interferometry, Master’s thesis, Faculty of Geode-
tic Engineering, Delft University of Technology.
Sandwell, D. T. and Price, E. J. (1998), Phase gradient approach to stacking interferograms,
Journal of Geophysical Research, 103(B12):30183–30204.
Sandwell, D. T. and Sichoix, L. (2000), Topographic Phase Recovery from Stacked ERS
Interferometry and a Low Resolution Digital Elevation Model, Journal of Geophysical
Research, 105(B12):28211–28222.
Sandwell, D. T., Sichoix, L., Agnew, D., Bock, Y. and Minster, J.-B. (2000), Near real-
time radar interferometry of the Mw 7.1 Hector Mine Earthquake, Geophysical Research
Letters, 27(19):3101–3104.
Scharroo, R. and Visser, P. (1998), Precise orbit determination and gravity field improve-
ment for the ERS satellites, Journal of Geophysical Research, 103(C4):8113–8127.
Schmetz, J., Menzel, W., Velden, C., Wu, X., van de Berg, L., Nieman, S., Hayden,
C., Holmlund, K. and Geijo, C. (1995), Monthly mean Large-scale analyses of upper-
tropospheric humidity and wind field divergence derived from three Geostationary Satel-
lites, Bulletin of the American Meteorological Society, 76(9):1578–1584.
Schwäbisch, M. (1995a), Die SAR-Interferometrie zur Erzeugung digitaler Geländemodelle,
Ph.D. thesis, Stuttgart University.
Schwäbisch, M. (1995b), Die SAR-Interferometrie zur Erzeugung digitaler Geländemodelle,
Forschungsbericht 95-25, Deutsche Forschungsanstalt fur Luft- un Raumfahrt, Oberpfaf-
fenhofen.
Schwäbisch, M. and Geudtner, D. (1995), Improvement of Phase and Coherence Map Qual-
ity Using Azimuth Prefiltering: Examples from ERS-1 and X-SAR, in: International
Geoscience and Remote Sensing Symposium, Florence, Italy, 10–14 July 1995, pp. 205–
207.
Segall, P. (1985), Stress and Subsidence Resulting From Subsurface Fluid Withdrawal in
the Epicentral Region of the 1983 Coalinga Earthquake, Journal of Geophysical Research,
90(B8):6801–6816.
Segall, P. and Davis, J. L. (1997), GPS applications for geodynamics and Earthquake
studies, Annu. Rev. Earth Planet. Sci., 25:301–336.
Seymour, M. S. and Cumming, I. G. (1994), Maximum Likelyhood Estimation For SAR
Interferometry, in: International Geoscience and Remote Sensing Symposium, Pasadena,
CA, USA, 8–12 August 1994, pp. 2272–2275.
Shapiro, I. I., Zisk, S. H., Rogers, A. E. E., Slade, M. A. and Thompson, T. W. (1972),
Lunar Topography: Global Determination by Radar, Science, 178(4064):939–948.
Sigmundsson, F., Durand, P. and Massonnet, D. (1999). Opening of an eruptive fissure
and seaward displacement at Piton de la Fournaise volcano measured by RADARSAT
satellite radar interferometry, Geophysical Research Letters, 26(5):533.
294 Bibliography

Sigmundsson, F., Vadon, H. and Massonnet, D. (1997), Readjustment of the Krafla spread-
ing segment to crustal rifting measured by Satellite Radar Interferometry, Geophysical
Research Letters, 24(15): 1843–1846.
Simpson, J. E. (1994), Sea Breeze and Local Wind, Cambridge University Press, Cambridge.
Skolnik, M. I. (1962), Introduction to Radar Systems, McGraw-Hill Kogakusha, Ltd.,
Tokyo.
Small, D., Werner, C. and Nüesch, D. (1993), Baseline Modelling for ERS-1 SAR Inter-
ferometry, in: International Geoscience and Remote Sensing Symposium, Tokyo, Japan,
18–21 August 1993, pp. 1204–1206.
Smith, Jr., E. K. and Weintraub, S. (1953), The Constants in the Equation for Atmospheric
Refractive Index at Radio Frequencies, Proceedings of the I.R.E., 41:1035–1037.
Smith, F. G., Little, C. G. and Lovell, A. C. B. (1950), Origin of the fluctuations in the
intensity of radio waves from galactic sources, Nature, 165:422–424.
Solheim, F. S., Vivekanandan, J., Ware, R. H. and Rocken, C. (1997), Propagation Delays
Induced in GPS Signals by Dry Air, Water Vapor, Hydrometeors and Other Particulates,
draft for J. of Geoph. Res.
Spoelstra, T. A. T. (1997), Personal communication.
Spoelstra, T. A. T. and Yi-Pei, Y. (1995), Ionospheric scintillation observations with radio
interferometry, Journal of Atmospheric and Terrestrial Physics, 57(1):85–97.
SPOT image company (1995), Presentation of SPOT Image Company, (Last modified at 2
Oct 1995), <http://www.orstom.fr/hapex/spot/spotimg.htm>.
Stein, R. S. (1999), The role of stress transfer in earthquake occurence, Nature, 402:605–
609.
Stein, R. S., Barka, A. A. and Dieterich, J. H. (1997), Progressive failure on the North
Anatolian fault since 1939 by earthquake stress triggering, Geophysical Journal Interna-
tional, 128:594–604.
Stoffelen, A. C. M. (1998), Scatterometry, Ph.D. thesis, Utrecht University.
Stolk, R., Hanssen, R., van der Marel, H. and Ambrosius, B. (1997), Troposferische sig-
naalvertraging; een vergelijking tussen GPS en SAR interferometrie, GPS Nieuwsbrief,
2:3–7, In Dutch.
Stolk, R. M. (1997), GPS and SAR interferometry, a comparative analysis of estimating
tropospheric propagation delays: The GISARE GPS Experiment, Master’s thesis, Delft
University of Technology, Delft.
Stotskii, A. A. (1973), Concerning the Fluctuation Characteristics of the Earth’s Atmo-
sphere, Radiophysics and quantum electronics, 16:620–622.
Stramondo, S., Tesauro, M., Briole, P., Sansosti, E., Salvi, S., Lanari, R., Anzidei, M.,
Baldi, P., Fornaro, G., Avallone, A., Buongiorno, M. F. and Boschi, G. F. E. (1999), The
September 26, 1997 Colfiorito, Italy, earthquakes: modeled coseismic surface displace-
ment from SAR interferometry and GPS, Geophysical Research Letters, 26(7):883–886.
Strang, G. (1986), A Proposal for Toeplitz Matrix Calculations, Studies in applied mathe-
matics, 74:171–176.
Strohmer, T. (1997), Computationally Attractive Reconstruction of Bandlimited Im-
ages from Irregular Samples, IEEE Transactions on Geoscience and Remote Sensing,
6(4):540–548.
Stull, R. B. (1995), Meteorology Today For Scientists and Engineers, West Publishing,
St. Paul, Minneapolis.
Susskind, Rosenfield, J. J., Reuter, D. and Chahine, M. T. (1984), Remote sensing of
weather and climate parameters from HIRS2/MSU on TIROS-N, Journal of Geophysical
Research, 89D:4677-4697.
Swart, L. M. T. (2000), Spectral filtering and oversampling for radar interferometry, Mas-
 Bibliography 295

ter’s thesis, Faculty of Geodetic Engineering, Delft University of Technology.

Tarayre, H. (1996), Extraction de Modeles Numeriques de Terrain par Interferometrie
Radar Satellitaire: Algorithmie et Artefacts Atmospheriques, Ph.D. thesis, Institut Na-
tional Polytechnique de Toulouse, Toulouse.
Tarayre, H. and Massonnet, D. (1996), Atmospheric propagation heterogeneities revealed
by ERS-1 interferometry, Geophysical Research Letters, 23(9):989-992.
Tatarski, V. I. (1961), Wave propagation in a turbulent medium, McGraw-Hill, New York,
Translated from Russian.
Taverne, B. G. (1993), Beginselen en voorbeelden van regelgeving en beleid ten aanzien van
de opsporing en winning van aardolie en aardgas, Delft University Press, Delft, in Dutch.
Taylor, G. I. (1938), The spectrum of turbulence, Proceedings of the Royal Society London,
Series A, CLXIV:476–490.
TDN/MD (1997), TopHoogteMD Digital Elevation Model, Topografische Dienst Nederland
and Meetkundige Dienst Rijkswaterstaat.
Tesauro, M., Berardino, P., Lanari, R., Sansosti, E., Fornaro, G. and Franceschetti, G.
(2000), Urban subsidence inside the city of Napoli (Italy) observed by satellite radar
interferometry, Geophysical Research Letters, 27(13): 1961.
Teunissen, P. J. G. (2000), Adjustment theory; an introduction, Delft University Press,
Delft, 1st edn.
Thatcher, W. and Massonnet, D. (1996), Crustal deformation at Long Valley Caldera.
eastern California, 1992-1996 inferred from satellite radar interferometry, Geophysical
Research Letters, 24(20):2519–2522.
Thayer, G. D. (1974), An Improved Equation for the Radio Refractive Index of Air, Radio
Science, 9:803–807.
Thiel, K.-H., Wu, X. and Hartl, P. (1997), ERS-tandem-interferometric obsertvation of
volcanic activities in Iceland, in: Third ERS Symposium—Space at the Service of our
Environment, Florence, Italy, 17–21 March 1997, ESA SP-414, pp. 475–480.
Thompson, A. R., Moran, J. M. and Swenson, G. W. (1986), Interferometry and Synthesis
in Radio Astronomy, Wiley-Interscience, New York.
Tough, R. J. A., Blacknell, D. and Quegan, S. (1995), A statistical description of polari-
metric and interferometric synthetic aperture radar, Proceeedings of the Royal Society
London A, 449:567–589.
Touzi, R. and Lopes, A. (1996), Statistics of the Stokes Parameters and of the Complex
Coherence Parameters in One-Look and Multilook Speckle Fields, IEEE Transactions
on Geoscience and Remote Sensing, 34(2):519–531.
Touzi, R., Lopes, A., Bruniquel, J. and Vachon, P. W. (1999), Coherence Estimation for
SAR Imagery, IEEE Transactions on Geoscience and Remote Sensing, 1(37): 135–149.
Touzi, R., Lopes, A. and Vachon, P. W. (1996), Estimation of the coherence function for
interferometric SAR applications, in: EUSAR’96, Königswinter, Germany, pp. 241–244.
Treuhaft, R. N. and Lanyi, G. E. (1987), The effect of the dynamic wet troposphere on
radio interferometric measurements, Radio Science, 22(2):251–265.
Tubbs, B. (1997), Astronomical Optical Interferometry, A Literature Review, (Updated
2000), <http://www.cus.cam.ac.uk/~rnt20/interferometry/ast_opt_int/pagel.html>.
Turcotte, D. L. (1997), Fractals and chaos in geology and geophysics, Cambridge University
Press, Cambridge, 2nd edn.
Ulaby, F. T., Moore, R. K. and Fung, A. K. (1982), Microwave remote sensing: active
and passive. Vol. 2. Radar remote sensing and surface scattering and emission theory,
Addison-Wesley, Reading.
Usai, S. (1997), The use of man-made features for long time scale INSAR, in: International
Geoscience and Remote Sensing Symposium, Singapore, 3–8 Aug 1997, pp. 1542–1544.
296 Bibliography

Usai, S., Gaudio, C. D., Borgstrom, S. and Achilli, V. (1999), Monitoring Terrain Deforma-
tions at Phlegrean Fields with SAR Interferometry, in: Second International Workshop
on ERS SAR Interferometry, ‘FRINGE99’, Liège, Belgium, 10–12 Nov 1999, pp. 1–5,
ESA.
Usai, S. and Hanssen, R. (1997), Long time scale INSAR by means of high coherence
features, in: Third ERS Symposium—Space at the Service of our Environment, Florence,
Italy, 17–21 March 1997, pp. 225–228.
Usai, S. and Klees, R. (1998), On the Feasibility of Long Time Scale INSAR, in: Inter-
national Geoscience and Remote Sensing Symposium, Seattle, Washington, USA, 6–10
July 1998, pp. 2448–2450.
Usai, S. and Klees, R. (1999a), On the interferometric characteristics of anthropogenic
features, in: International Geoscience and Remote Sensing Symposium, Hamburg, Ger-
many, 28 June–2 July 1999.
Usai, S. and Klees, R. (1999b), SAR Interferometry On Very Long Time Scale: A Study
of the Interferometric Characteristics Of Man-Made Features, IEEE Transactions on
Geoscience and Remote Sensing, 37(4):2118–2123.
Visser, P. N. A. M., Scharroo, R., Floberghagen, R. and Ambrosius, B. A. C. (1997),
Impact of PRARE on ERS-2 orbit determination, in: Proceedings of 12th International
Symposium on ”Space Flight Dynamics”, Darmstadt, Germany, 2-6 June 1997, pp. 115–
120, SP3-403.
Voss, R. F. (1988), Fractals in nature: From characterization to simulation, in: Peitgen
and Saupe (1988), pp. 21–70.
Wackernagel, H. (1995), Multivariate Geostatistics: An Introduction with Applications,
Springer-Verlag, Berlin.
Wadge, G., Scheuchl, B. and Stevens, N. F. (1999), Spaceborne radar measurements of
the eruption of Soufriere Hills Volcano, Montserrat during 1996-99, Submitted to J.
Volcanology and Geothermal Research.
Weckwerth, T. M., Wilson, J. W. and Wakimoto, R. M. (1996), Thermodynamic Variability
within the Convective Boundary Layer Due to Horizontal Convective Rolls, Monthly
Weather Review, 124(5):769–784.
Weckwerth, T. M., Wilson, J. W., Wakimoto, R. M. and Crook, N. A. (1997), Horizontal
Convective Rolls: Determining the environmental conditions supporting their existence
and characteristics, Monthly Weather Review, 125(4):505–526.
Weckwerth, T. M., Wulfmeyer, V., Wakimoto, R. M., Hardesty, R. M., Wilson, J. W.
and Banta, R. M. (1999), NCAR/NOAA Lower-Tropospheric Water Vapor Workshop,
Bulletin of the American Meteorological Society, 80(11):2339–2357.
Wegmüler, U., Werner, C. L., Nüesch, D. and Borgeaud, M. (1995), Land-Surface Analysis
Using ERS-1 SAR Interferometry, ESA Bulletin, 81:30–37.
Wegmüller, U. and Werner, C. (1997), Retrieval of Vegetation Parameters with SAR In-
terferometry, IEEE Transactions on Geoscience and Remote Sensing, 35(1):18–24.
Wegmüller, U. and Werner, C. L. (1995), SAR Interferometric Signatures of Forest, IEEE
Transactions on Geoscience and Remote Sensing, 33(5): 1153–1161.
Wegmüller, U., Werner, C. L., Nüesch, D. and Borgeaud, M. (1995), Forest Mapping Using
ERS Repeat-Pass SAR Interferometry, Earth Observation Quarterly, 49.
Weldon, R. B. and Holmes, S. J. (1991), Water Vapor Imagery: Interpretation and Applica-
tions to Weather Analysis and Forecasting, Tech. Rep. NESDIS 57, NOAA, Washington.
Werner, C. L., Hensley, S., Goldstein, R. M., Rosen, P. A. and Zebker, H. A. (1993),
Techniques and applications of SAR interferometry for ERS-1: Topographic mapping,
change detection and slope measurement, in: First ERS-1 Symposium—Space at the
Service of our Environment, Cannes, France, 4–6 November 1992, ESA SP-359, pp.
 Bibliography 297

205–210.
Werner, M. (1998), Shuttle Radar Topography Mission (SRTM). The X-Band SAR Inter-
ferometer, in: 28th EuMC workshop proceedings, EuMW Amsterdam, 9 October 1998.
Wessel, P. and Smith, W. H. F. (1998), New, Improved Version of Generic Mapping Tools
Released, EOS Transactions, AGU, 79(47):579.
Weydahl, D. J. (1991), Change detection in SAR images, in: International Geoscience and
Remote Sensing Symposium, Espoo, Finland, 3–6 June 1991, pp. 1421–1424.
Wicks, Jr, C., Thatcher, W. and Dzurisin, D. (1998), Migration of Fluids Beneath Yellow-
stone Caldera Inferred from Satellite Radar Interferometry, Science, 282:458–462.
Wiley, C. A. (1954), Pulsed Doppler Radar Methods and Apparatus, United States Patent,
No. 3196436.
Williams, C. A. and Wadge, G. (1998), The effects of topography on magma chamber
deformation models: Application to Mt. Etna and radar interferometry, Geophysical
Research Letters, 25(10):1549–1552.
Williams, S., Bock, Y. and Fang, P. (1998), Integrated satellite interferometry: Tropo-
spheric noise, GPS estimates and implications for interferometric synthetic aperture
radar products, Journal of Geophysical Research, 103(B11):27,051–27,067.
Wu, C., Barkan, B., Huneycutt, B., Leans, C. and Pang, S. (1981), An introductoin to
the interim digital SAR processor and the characteristics of the associated Seasat SAR
imagery, JPL Publications, 81-26, p. 123.
Wu, X. (1996), Anwendung der Radarinterferometrie zur Bestimmung der Topographie
und der Geschwindigkeitsfelder der Eisoberflächen antarktischer Gebiete, Ph.D. thesis,
Universität Stuttgart.
Yocky, D. A. and Johnson, B. F. (1998), Repeat-Pass Dual-Antenna Synthetic Aperture
Radar Interferometric Change-Detection Post-Processing, Photogrammetric Engineering
and Remote Sensing, pp. 425–429.
Zebker, H. (1996), Imaging Radar and Applications, Class notes EE355.
Zebker, H., Rosen, P., Hensley, S. and Mouginis-Mark, P. (1996), Analysis of active lava
flows on Kilauea volcano, Hawaii, using SIR-C radar correlation measurements, Geology,
24:495–498.
Zebker, H. A., Farr, T. G., Salazar, R. P. and Dixon, T. H. (1994a), Mapping the World’s
Topography Using Radar Interferometry: The TOPSAT Mission, Proceedings of the
IEEE, 82(12):1774–1786.
Zebker, H. A. and Goldstein, R. M. (1986), Topographic Mapping From Interferometric
Synthetic Aperture Radar Observations, Journal of Geophysical Research, 91(B5):4993–
4999.
Zebker, H. A. and Lu, Y. (1998), Phase unwrapping algorithms for radar interferometry:
residue-cut least-squares, and synthesis algorithms, Journal of the Optical Society of
America A., 15(3):586–598.
Zebker, H. A., Madsen, S. N., Martin, J., Wheeler, K. B., Miller, T., Lou, Y., Alberti,
G., Vetrella, S. and Cucci, A. (1992). The TOPSAR Interferometric Radar Topographic
Mapping Instrument, IEEE Transactions on Geoscience and Remote Sensing, 30(5):933–
940.
Zebker, H. A., Rosen, P. A., Goldstein, R. M., Gabriel, A. and Werner, C. L. (1994b), On
the derivation of coseismic displacement fields using differential radar interferometry:
The Landers earthquake, Journal of Geophysical Research, 99(B10): 19617–19634.
Zebker, H. A., Rosen, P. A. and Hensley, S. (1997), Atmospheric effects in interferometric
synthetic aperture radar surface deformation and topographic maps, Journal of Geo-
physical Research, 102(B4):7547–7563.
Zebker, H. A. and Villasenor, J. (1992), Decorrelation in interferometric radar echoes, IEEE
298 Bibliography

Transactions on Geoscience and Remote Sensing, 30(5):950–959.

Zebker, H. A., Werner, C. L., Rosen, P. A. and Hensley, S. (1994c), Accuracy of Topographic
Maps Derived from ERS-1 Interferometric Data, IEEE Transactions on Geoscience and
Remote Sensing, 32(4):823–836.
Zisk, S. H. (1972a), Lunar Topography: First Radar-interferometer Measurements of the
Alphonsus-Arzachel Region, Science, 178(4064):977–980.
Zisk, S. H. (1972b), A New Earth-based Radar Technique for the Measurement of Lunar
Topography, Moon, 4:296–306.
 About the Author

Ramon Hanssen (1968) studied aerospace engineering (1987-1989) and geodetic engi-
neering (1989-1993) at Delft University of Technology (M.Sc. 1993). His graduation
work focused on the use of the Radon transformation for the analysis of sparsely
sampled data, with applications in satellite altimetry and imaging tomography. In
1994, he worked as a researcher on potential field (gravity and aeromagnetics) data
inversion at the International Institute for Aerospace Surveys and Earth Sciences
(ITC). In 1995, he started his PhD research at the Delft Institute for Earth-Oriented
Space research (DEOS) of Delft University of Technology on the geodetic analysis
of repeat-pass spaceborne radar interferometry, with emphasis on error propaga-
tion. During this research, he worked as a visiting scholar at Stuttgart University
(1996), at the German Aerospace Center (DLR, 1997) and at Stanford University
(1997-1998) on a Fulbright research fellowship. In 1997, he received an ESA grant to
study the effect of atmospheric heterogeneities for ERS-1/2 radar interferometry, a
project performed in close collaboration with the Royal Netherlands Meteorological
Institute (KNMI). At present, he is employed as an assistant-professor in the field of
geostatistics at the Faculty of Civil Engineering and Geosciences at DUT. His cur-
rent research interests are the geodetic analysis of two-dimensional survey data, the
influence of atmospheric delay on space-geodetic techniques, and the mathematical
modeling and physical interpretation of deformation processes.
 Index
Absolute phase, 63 variability, 66
Accuracy, 62, 64 variation, 263
Adjustment, 5, 62 Autocorrelation, 82
Aliasing, 270 AVHRR, 44, 219, 239, 260, 260
Alignment, 36, 77 resolution, 243
All-weather capabilities, 71 Azimuth
Almaz, 11, 17 ambiguities, 31
ALOS, 17 bandwidth, 25, 32, 87, 103
Ambiguity beamwidth, 30
height, 37 compression, see Azimuth focusing
interferogram, 71 filtering, 42, 45, 50–52, 103
number, 64 focusing, 25, 32
Antenna line number, 46
gain, 101 look direction, 163, 192
Approximation error, 67, 68 offset vector, 46, 214
Ascending orbit, 70 oversampling, 108
Atmosphere, 5, 197, 198 position, 59
frozen, 242, 265, 269 reference function, 32
middle, 198 resolution, 31, 32
neutral, 198 start time, 59
superposition, 71 time, 34
Atmospheric
correction, 130, 262, 263 Baseline, 66, 114
delay, 2, 3, 64, 73 critical, 16, 40, 43, 102, 103
dynamics, 7, 197 effective, see Baseline perpendicular
mapping, 61, 70 error, 129
model, 145, 250 parallel, 129, 130
observations, 139 perpendicular, 124, 128–130, 249
physics, 2 inventory plot, 43, 44
reduction, 64 length, 63, 67
refractivity, 2, 5 normal, see Baseline perpendicular
regimes, 142 orientation, 63
signal, 5, 64, 65, 71, 75, 131–133, 136, 140, parallel, 45, 53, 115, 116, 119, 120
142–144, 198, 248, 258 perpendicular, 22, 35, 37, 38–40, 43, 44,
amplification, 76 49, 53, 57–59, 64, 68, 77, 78, 115,
amplitude and phase, 220–226 124–126, 129, 130
calm weather, 146 deformation, 43
cold front, 216 geometric decorrelation, 103
convection, 221 integer-scaled, 77
elimination, 5 sign, 119n
identification, 187 topography, 43
modeling, 5 ratio, 75, 76
precipitating cumulonimbus, 215 scaling, 75, 76
precipitation, 221 sign of, 116
reduction, 76 spatial, 35, 43
rolls, 218, 225, 227 temporal, 35, 43, 44, 64, 77, 154, 164, 178,
stochastic, 76 183, 249, 252
stratification, 148 deformation, 44
summation, 137 topography, 43
suppressing, 73 zero, 38, 75
thunderstorm, 146 Baseline decorrelation, see Geometric decorre-
turbulence, 130 lation
types of, 131 Bessel function, 93
weighting, 76, 77 Bistatic radar, 9
signal delay, 6, 255 Block Toeplitz-Toeplitz Block, 157
simulation, 147 matrix, 158, 159, 252
stratification, 148–154 Bootstrap method, 250

301
302 INDEX

Boundary layer, 243, 269 repeated, 57

convection, 218 review, 47
depth, 219, 225 shift, 46
effective height, 242 techniques, 46
frozen, 242 Corner reflector, 7, 162, 164–169
moisture distribution, 220, 245 detection, 168
rolls, 218, 220, 221, 225, 228, 245 experiments, 163–173
turbulence, 242 height, 165
Bragg scattering, 221 intensity, 169, 170
Brownian motion, 84 leveling, 167
fractional, 86, 88 location, 166, 168
BTTB, see Block Toeplitz-Toeplitz Block movement, 6, 21, 40, 164–166, 168–172
phase, 165, 168, 169
C-band, 199, 212 phasor, 169
Calibration, 72 reference, 169
Cerro Prieto, 175 scattering, 156
deformation mechanisms, 186 sidelobes, 168
forward modeling, 186 spatial scales, 165
interferograms, 179–181 statistics, 169
Chirp, 22, 28, 29, 32 Correlation, 66
compression, 27, 28, 29 total. 99
replica, 28 Cosmos, 11, 17
scaling, 24 Covariance, 3
slope, 27 function, 6, 82, 144
waveform, 27 and power spectrum, 137
Circular matrix, 62
Gaussian PDF, 90 Cramér-Rao bound, 95
Gaussian signal, 106 Critical baseline
Gaussian statistics, 91 seeBaseline critical, 16
Gaussian variable, 89 Crosswind direction, 223
CLARA 96, 237–244
Clausius-Clapeyron equation, 204 Data
Clausius-Mossotti, 206 acquisition, 5–7
Climate studies, 197, 198 adjustment, 6
Clock error, 228 availability, 3, 4
Cloud interpretation, 5
cover, 260 processing, 4–7
development, 205 quality, 4
formation, 198 storage, 3
liquid water content, 206 Decorrelation, 3, 6, 39
signal propagation, 205 Doppler centroid, 98, 103, 248
CMOD4, 221, 223 geometric, 18, 21, 98, 102, 102, 103, 104,
Cofactor matrix, 61 115, 154, 155, 179, 248
Coherence, 65, 70, 92, 94, 95, 96, 97, 106, 201 filtering, 102
bias, 97, 174 model, 22
complex, 96 processing induced, 81, 98
estimation, 96, 97, 173 sources, 98
PDF, 96 surface, 18
phase-corrected, 97 temporal, 2, 13, 16, 19, 21, 39, 41, 43, 56,
thermal noise, 102 66, 76, 81, 98, 103, 104, 109, 140,
total, 108, 111 154–156, 161, 164, 173, 174, 181,
Coherent 182, 185, 189, 192, 194, 196, 213,
imaging, 35 245
scatterers, 7 analysis, 104
Cold front, 222, 268 characteristics, 183
Conclusions, 247 indications, 248
Convection, 198 models, 104
Coregistration, 19, 30, 45, 51, 110, 248 prediction, 248
accuracy, 46, 110, 111 thermal, 98, 99, 155, 156
along-track, 121 total, 35
azimuth offset, 214 volume, 22, 98, 104
coarse, 45, 46 Deformation, 5, 64, 161
cross-correlation, 46 components, 69, 163
decorrelation, 98, 110, 111, 155 mapping, 61, 69
azimuth, 110 monitoring, 7
range, 110 pair, 75, 76
error, 110, 112 sensitivity, 70
fine, 45, 46 studies, 3, 64
 INDEX 303

vector, 163 Izmit, 12, 162, 190, 190, 191, 194

Delay Landers, 15
absolute-temporal, 265 risk, 190, 194
clouds, 206 Effective baseline, see Baseline perpendicular
dry, 255 Electromagnetic properties, 70
GPS and InSAR, 261 Electron density, 71, 202
gradients, 7, 202 Elevation
hydrostatic, 71, 203, 204, 207, 241, 242, angle, 257, 258, 263
244, 245, 255, 265 angle,cut-off, 259
accuracy, 203 model, 3, 201
ionospheric, 71, 209–215, 255 Ellipsoid, 52
lateral variation, 201, 204 Envisat, 17
liquid, 71 Equation of state, 202, 207
moisture contribution, 204 Ergodicity, 85
neutral, 255 Error
relative-spatial, 265 analysis, 6
sensitivity, 204 atmospheric, 7
GPS, 268 clock, 228, 257
InSAR, 268, 269 orbit, 7, 259
slant, 265 propagation, 5–8, 154, 248
spatial scale, 265 sources, 1–3, 5, 8, 81, 248
temperature contribution, 204 ERS, 3, 8, 12, 12, 16, 27–29, 40, 43, 46, 50, 70
time-series, 256 Euclidean geometry, 6
tropospheric, 255, 258, 259, 261, 267 Euler dilogarithm, 95
total, 71n Expectation, 82
variation estimate, 265
wet, 71, 203, 204, 255, 259, 261, 265 Far field approximation, 36, 67, 68, 70
Descending orbit, 70 Far-range, 24, 59, 116, 118–120, 124, 190
Design matrix, 61, 62, 65, 66, 73 FFA, see Far field approximation
partial differential, 75 Filtering, 6
Deskew, 30, 32 a posteriori, 53
Deterministic approach, 5 a priori, 47
Dielectric constant, 99n spectral shift, 53
Differential interferometry, 161, 265 Focused spot scanning, 26
four-pass, 76 Forecasting, 197, 251
three-pass, 74 numerical, 197, 198
Dipole moment, 132, 206 precipitation, 197
Dirichlet, 106n Foreshortening, 18, 32, 33, 54, 60
Dispersion, 65, 242 Four-pass interferometry, 73, 77
Dispersive delay, 207, 209, 255 Fourier transform, 274
Displacement vector, 69, 162 Fractal
Doppler, 23 dimension, 85, 86, 132, 141–143, 146
average, 103 and power spectrum, 141
bandwidth, 30 mdels, 88
ERS, 31 self-affine, 86
centroid, 23, 25, 29, 30, 31, 42, 47, 53, 59, self-similar, 86
103 simulation, 86
decorrelation, 98, 103 statistics, 86
estimation, 51 Fringe, 13, 14
frequency, 30, 31, 45, 47, 50, 51, 103 Functional model, 5, 61, 72, 248
processing, 32 atmosphere, 70
frequency, 30, 214 deformation, 69
textbf, 29 four-pass, 76
iso-, 14 generic, 65, 73
range-, 14, 25, 32 linear combinations, 77
spectrum, 30 single-pass, 66
steering, 103 stacks, 72
zero, 23, 30, 30, 42, 47, 49, 51, 53 three-pass, 74
frequency, 47 topography, 66
processing, 30, 32
zero-, 29 Gamma function, 93
Droplets, 198, 204, 205 Gauss-Markoff, 6, 7, 61
Generic functional model, 65
Early azimuth, 24 Geometric decorrelation, see Decorrelation
Earthquake, 12, 18–20, 40, 44, 177, 190, 192– Geometric range, 66, 70
194 Geophysical
deformation, 196 parameters, 6
history, 190 Geothermal field, 175
304 INDEX

GISARE, 165, 256 phase gradients, 125

-GPS, 258 phase noise, 98
test area, 166 processing, 47, 104
Global reference systems, 1 quality, 4–6, 81, 103, 105, 248
GOES, 197 residual fringes, 123, 124, 126, 128
GPS, 2, 5–7, 197, 255, 258, 260 scaling, 77, 81
and InSAR, 255 series, 131
comparison, 267 simulation, 108
cone, 259, 269 SNR, 115
delay sensitivity, 269 spikes, 109
initialization, 250 stacking, 61, 62n, 103, 131
network, 255, 264 stacks, 64, 72
sampling, 269 topographic, 15, 38, 57, 64, 75, 201
single station, 264 unwrapped, 55
slant delay, 200 vectorization, 63
spatial network, 256, 264 wrapped, 77, 78
time-series, 255, 259, 264, 265, 269, 270 Interferometric
tropospheric delay, 261 phase contributions, 199
zenith delay, 200 phase gradient, 62n
Gravity, 203, 212 processing, 5, 6, 9
waves, 212, 220, 260 Interferometric Radar Meteorology, 65, 197, 247
Groningen, 173 Interferometric radar meteorology, 250
Ground range, 35 Interferometry
Ground truth, 4 additive, 13
Gust front, 217, 221 differential, 74, 76
four-pass, 74
Hankel’s function, 93 limitations, 200
Hausdorff measure, 86 multi-baseline , 73
Height ambiguity, 37, 57, 60, 78, 115, 126, 165, multiplicative, 13
190 three-pass, 73
Height error, 158 Interpolation, 248
Hermitian product, 61, 92 errors, 104
Heterogeneity kernels, 106
ionospheric, 8 repeated, 57
tropospheric, 8 Intrinsic
Homogeneity, 85 functions, 83, 85, 86, 132
Humidity observations, 220 hypothesis, 84
Hydrostatic delay, 203 Ionosphere, 197, 198
Hypergeometric function, 94 Ionospheric
Hypothesis testing, 62 delay, 209–215
refractivity, 202
IRM, see Interferometric Radar Meteorology
Ideal gas law, 203 Isotropy, 85
Incidence angle, 70, 201 Iteration, 68
variable, 70 Izmit, 190–194
Inertial subrange, 133, 138 analysis, 192
InSAR, 3, 4, 9 interferogram, 190, 191
Integer ambiguity number, 64
Interferogram, 15, 200 JERS, 12, 17, 43
atmospheric, 128 Johnny Appleseed, 183n
atmospheric influence, 131 strategy, 184
bandwidth, 52 Joint stochastic model, 154–159
cascade, 71
complex, 36, 51
decorrelation, 105 Kolmogorov power spectrum, 134
deformation, 15, 38, 57, 75, 128 Kolmogorov turbulence, 133, 137
difference, 78 theory, 133, 147
differential, 15, 38, 57, 75, 201, 267
four-pass, 76 L-band, 212
three-pass, 76 Lacrosse, 17
filtering, 54 Land subsidence, 162–164, 178
focused, 62 Cerro Prieto, 175–189
formation, 48, 51, 53, 116 Groningen, 173–174
integer scaling, 77 Late azimuth, 24
interpretation, 130 Latent heat, 204
linear combination, 77 Layover, 18, 33, 42, 54, 56, 59, 60
linear integer combination, 57, 77, 79 Leveling, 6, 166, 167, 177
multiple, 73 Limitations, 5
noise, 52 Line of sight, 2
 INDEX 305

Linear transfer functions, 6 dispersion of, 65, 81, 82, 87

Liquid water equations, 7, 61, 75, 247
content, 71, 202, 207 meteorological, 6
density, 207 number of, 63, 64, 158
Look angle, 30, 34, 63, 201 period, 52
phase, 63
Magellan, 11, 17 principal, 54, 62
Mapping pseudo-, 249, 252
atmospheric, 7 space, 61
deformation, 7 space-geodetic, 2, 5, 7, 197
function, 201, 255, 257, 265 statistics, 81, 87
topographic, 7 vector, 62, 63, 65, 156
Mathematical model, 5 One-way path delay, 71
Mesoscale, 198 Orbit
surface winds, 220 accuracy, 8, 68, 114
wind, 245 determination, 114
Mesoscale Shallow Convection, 198, 226, 245 error correction, 125
Mesosphere, 198 error propagation, 128
Meteorological errors, 6, 113, 121–130
information, 5 Outline, 7
interpretation, 71, 198 Oversampling, 34, 46, 47, 50–52, 108, 270
observations, 6 decorrelation, 108
Meteorology, 6, 197 ratio, 34, 47, 107, 108
mesoscale, 197, 198
operational, 198 Pair-wise logic, 187
Meteosat, 197, 228, 233, 235, 236, 239, 260 Parallel ray approximation, see Far field ap-
and GPS, 232 proximation
and InSAR, 232 Parameter, 7, 61, 63, 64
brightness temperature, 235 dominant, 63
double visual, 241 estimation, 62
elevation angle, 234 reduction, 64
positioning accuracy, 233, 236 space, 61
precipitable water vapor, 228, 234 vector, 61, 74
radiometry, 226 Patch unwrapping, 57, 183
resolution, 243 PDF, see Probability Density Function
scan, 234 Permanent scatterers, 96
signal intensity, 230 Perpendicular baseline, see Baseline perpendic-
WV channel, 226, 230–233, 235 ular
Methodology, 6, 7 Phase
Michelson-Morley interferometer, 14 advancement, 70
Model ambiguity, 3, 6, 257
functional, 7 bias, 5
mathematical, 7 covariance, 248
observation equations, 75 delay, 70, 206
stochastic, 7 flattened, 74, 77, 78
Mogi source, 186, 188 geometric component, 200, 201
Moisture gradient, 62n, 201
lateral variation, 205
interpretation, 201
Moments
ML estimator, 93
first-order, 82
PDF, 93
second-order, 82
single-look, 94
Monostatic radar, 6, 9, 26
propagation component, 200, 201
Motivation, 1
reference, 119n
MSC, see Mesoscale Shallow Convection
Multi-baseline interferometry, 73
residue, 47, 55–57, 108, 112, 114, 184
Multilooking, 52, 87, 92
scattering component, 199, 200
Multipath, 257, 258
standard deviation, 95
unwrapping, 6, 42, 63, 64, 66, 111, 183,
201, 259
NDVI, see Normalized Differential Vegetation variance, 87, 94–96, 111, 147, 154–156,
Index 158, 248
Near-range, 24, 59, 115, 116, 118–120, 124, 190 velocity, 70
Noise Phase ambiguity
1/f, 86 integer, 36, 42, 55, 64, 68, 81, 111–113,
white, 86 158, 248, 249, 252, 259
Normalized Differential Vegetation Index, 44 Phasor, 36, 42, 51, 89, 169
Picture element, see Pixel
Objectives, 7 Pixel, 33, 34, 39, 46, 47, 55, 59, 111, 112
Observation, 61, 62, 64, 81, 87, 145 alignment, 46
306 INDEX

intensity, 91, 95 applications, 2, 7, 15, 16, 43. 70, 188,

interpolation, 168 219, 236, 244, 247
number, 113, 117 atmospheric model, 145, 146
phase, 51, 168 capabilities, 2
phase gradient, 185 differential, 15
position, 115, 121 difficulties, 54
range, 117 error sources, 81, 159, 255
reference, 158 experiments, 164
registration, 45 feasibility, 8, 19, 43, 161, 164, 194, 195
size, 33 for meteorology, 8, 198
spacing, 34 functional modal, 61
value, 237 geophysical parameters, 63
variance, 159 limitations, 5, 237
Planetary boundary layer, 198, 226 measurements, 1
PLW, see Precipitable Liquid Water observations, 220
Point scatterers, 95 parameter estimation, 87
Poroelastic, 186–188 potential, 237
Positive definite, 62, 83 stochastic model, 81, 147, 154
Posting, 33, 34, 34, 45, 46, 87, 111, 157, 242, subsidence, 188
259 validation, 226–237
Power consumption, 3 value of, 6
Power spectral density, see power spectrum system theory, 9
Power spectrum, 132, 271, 273, 274 wavelength, 2
and covariance function, 137 Radarsat, 13, 17, 43
and fractal dimension, 141 Radiation budget, 198
and structure function, 134 Radio wave propagation, 8, 70, 255
Power-law, 273 Radiometer, 197
Precipitable liquid water, 208 Radiosonde, 259, 260, 265
Precipitable water vapor, 199, 205–207, 207, Random
208, 215, 232, 238–240, 251 functions, 85
Pressure, 71, 199 walk, 86
dry air, 202 Range
gradients, 203 -azimuth coordinates, 32, 58
saturation, 204 -azimuth ratio, 52
surface, 203 bandwidth, 17, 43, 87, 101
total, 202 bin, 31, 32, 38, 59
water vapor, 202, 204, 205 compression, 25, 28, 29, 32
PRF, see Pulse repetition frequency deformation, 70, 179, 195, 252
Probability Density Function, 6, 91, 199, 248 difference, 26, 52, 67
exponential, 92 absolute, 68
joint, 90, 92, 129 approximate derivation, 67
marginal, 90, 91, 93, 129, 130 approximation error, 68
Rayleigh, 91, 92 exact derivation, 67
uniform, 89, 91, 92, 200 relative, 69
Processing errors, 5, 98 true, 68
Propagation, 2 direction, 27, 59
delay, 197 displacement, 170, 171, 186, 192
velocity, 70 enhancement factor, see time-bandwidth
Pulse product
compression, see Range compression error, 210, 211
effective length, 27 filtering, 47–50, 52
repetition frequency (PRF), 22, 26, 28, 30 geometric, 69
PWV, see Precipitable water vapor ground, 99
migration, 25, 31, 32, 32
modulation, 26
Quadrature demodulation, 28 of the covariance function, 83
offset vector, 53
Radar, 9 oversampling, 108
altimetry, 2, 3, 5 posting, 46
basic concepts, 9 reference function, 28, 29, 32
cross section, 99n, 156 reference phase, 117
footprint, 200 resolution, 16, 27, 43, 89, 110
frequency, 71, 202 sensitivity, 10
fundamentals, 7 slant, 99
history, 9 vector, 53, 59
interferogram, 4, 5, 128, 139, 215, 221, Rank defect, 64, 66
233, 247 Rank deficiency, 63, 72
interferometry, 2 Ray bending, 70
accuracy, 148, 154 Ray path, 70
 INDEX 307

Rayleigh distribution, 91 Satellite payload, 3

Real aperture radar, 26 Saturation pressure, 204
Recommendations, 251 Scale-invariance, 85
Rect function, 106n Scaling, see Scale-invariance
Recurrence relation, 275 ScanSAR, 11, 13, 245
Redundancy, 248 Scattering, 9
Reference distributed, 89, 91, 93, 158
body, 62, 63, 65, 67 Gaussian, 89
datum, 4 point, 89
elevation model, 64 Rayleigh, 89
phase, 38, 62, 63, 113 specular, 42n
systems, 1 Seasat, 11, 15, 17, 50
Reflectors, 77 Semi-variogram, 82
Refractive index, 70 Shadow, 18, 33, 59, 60
Refractivity, 7, 70, 71, 133, 197, 199, 200, 202 Shift theorem, 274
and temperature, 204 Sidelobe, 29, 42, 168
clear-air, 204 Signal delay, 199, 200, 205, 221, 255, 271
components, 72 rule of thumb, 71
decomposition, 71 total, 71
distribution, 72, 134, 199, 265 variability, 199
dry, 202 Signal interpretation, 6
hydrostatic, 71, 203 Signal-to-noise, 4, 25, 96, 98, 100
integrated, 8, 70, 71, 201, 202 Sill, 83, 136
ionospheric, 71, 202 Simulations, 6
lateral variability, 72 Single-Look Complex, 42
liquid, 71, 202 Single-pass, 2, 66
mean, 72 SIR-A, 11, 11, 17
sensitivity, 204, 205, 208 SIR-B, 11, 17
total, 204 SIR-C, 12, 17
vertical variation, 72 Skewed processing, 32
wet, 71, 202, 205 Slant
Relative humidity, 204, 205, 217 deformation, 69
Reliability, 8, 62 delay, 66, 70, 70
Relief, 76 Slant delay, 202
Remote sensing, 1 Slant range, 33, 35, 70
optical, 5 SLC, see Single-Look-Complex
Repeat-pass, 2, 66 Slope-adaptive filtering, 102
Resampling, 36, 73 Slow deformation processes, 73
Research Smoothness criterion, 55
limitations, 6 Snell’s law, 42n
methodology, 6 SNR, see Signal-to-noise
objectives, 4 Spatial correlation, 156
questions, 7, 247 Speckle, 77, 89, 91, 174
Resel, see Resolution element or Resolution cell tracking, 19
Residue, see Phase residue Spectral
Resolution, 2, 10, 27, 33, 33, 52, 71, 221 approximations, 76
azimuth, 11, 23, 25 density, see Power spectrum
focused, 31 representations, 6
real-aperture, 26 shift, 48
cell, 5, 32, 33, 33, 34–37, 42. 52, 66, 68, azimuth, 50
69, 87, 154, 199 range, 49
correlation, 66, 81, 87 Specular reflection, 42, 42n
PDF, 89, 91 Spotlight SAR, 11
element, 15 Squint
range, 26, 27, 35, 37, 50 angle, 29–31, 50, 53, 59
accuracy, 35 variation, 50, 51
effective, 27 zero-, 32
Squint mode SAR, 11
improvement, 28
SRTM, 13, 17, 18, 245
spatial, 6, 8 Stacking, 19, 48, 61, 62n, 64, 72, 73, 95, 103,
180–183, 252
Saastamoinen model, 203, 265 Stationarity, 83, 85, 137
Saastamoinen-Baby model, 153 quasi, 84, 85
SAR, 9 second-order, 84, 85
focusing, 6 spectral exponent, 86
image mode, 22 strict, 84, 86
Satellite missions Stationary increments, 84–86
future, 8 Statistical interpretation, 6
Satellite orbits, 3 Statistics
308 INDEX

multiple-point, 7 delay, 257

point, 6 Turbulence, 6, 130
single-point, 7 inner scale, 133
Stochastic outer scale, 133
approach, 5 Turbulent mixing, 131
model, 3, 5, 61, 64, 66 Two-thirds law, 133
Stochasticity, 62
Stratocumulus, 198 Unknowns, see parameters
Stratosphere, 198 Upwind direction, 223
Structure function, 6, 82, 132, 145, 237, 239,
242, 271–273 Variance, 5
and power spectrum, 134 factor, 61
of delay, 134, 271, 273 of random process, 274
of refractivity, 133, 271, 273 Variance-covariance matrix, 3, 6, 61, 65, 66, 87,
two-dimensional, 244 88, 132, 147, 154, 156–159
Subsidence Variogram, 82, see structure function
inversion, 233 Vectorization, 63
monitoring, 7 Vegetation, 66, 76
rate, 163 Vertical
Surface deformation, 70
deformation, 64, 66, 73, 199, 248 stratification, 131, 148
pressure, 203 Very Long Baseline Interferometry, see VLBI
Surplus value, 4 VLBI, 2, 3, 5, 7, 197
Swath, 2, 10, 13, 17, 24, 26, 40, 228, 265
overlap, 43
width, 2, 16, 26, 197, 221, 244, 245 Water vapor, 2, 197, 198, 207, 259
concentrations, 197
limit, 40
Synoptic density, 207
distribution, 197, 247
data, 139, 267 mapping, 198, 215–220
observations, 45, 245, 260, 266 maps, 197
Synthetic Aperture Radar, see SAR observed by InSAR, 270
partial pressure, 71
Tandem mission, 3, 12, 16, 140, 141, 162, 164, spatial variation, 199
165, 179, 180, 182, 200, 245, 250, Wavefront, 134
256 Wavelength, 66
Taylors hypothesis, 242, 265 Weather radar, 260
Temperature, 71, 198, 199, 202, 204 Weight matrix, 62
lapse rate, 205 WGS84, 52, 68
observations, 220 Wiener-Khinchine theorem, 137
receiver noise, 100 Wind measurements, 220
Temporal baseline, see Baseline Wrapped phase, 62, 63
Temporal decorrelation, see Decorrelation Wrapping operator, 62
Thematic mapping, 21
Thermal X-band, 212
decorrelation, 98, 101 X-SAR, 12, 17
noise, 21, 40, 165, 248
coherence, 102 Yaw-steering, 23, 50, 103
power, 100
Thermoelastic, 186–188
Three-pass interferometry, 73 Zenith
Thunderstorm, 197 angle, 70, 258
initiation, 197 delay, 201, 202, 255–259, 261, 263–265,
Tie-point unwrapping, 184 267
Tie-points, 63 Zero-padding, 29, 51, 168
Time-bandwidth product, 27
Time-series, 73
Toeplitz, 157–159, 252
Topographic
height, 63, 64, 73, 199, 201, 248
information, 64
pair, 75–77
phase, 62
Topography, 5
estimation, 61, 66
Triangulation, 6
Troposphere, 6, 197, 198, 255
destabilization, 198
Tropospheric